awesome-egocentric-atlas

KO | EN

gitlite — search

#openai #nlp #awesome-list #ai #docker #claude-code #security #hacktoberfest #llm #codex

awesome-egocentric-atlas

(★ 9)

Curated datasets, benchmarks, models, and tools for egocentric AI, embodied intelligence, VLA, world models, robotics, and wearable vision.

파일 탐색기

ZIP 다운로드

- - add_resource.yml
  - config.yml
  - fix_or_flag.yml
- - freshness.yml
  - linkcheck.yml
  - pages.yml
  - sync-huggingface.yml
  - validate.yml
- PULL_REQUEST_TEMPLATE.md
- - adl-dataset.png
  - adl-dataset.webp
  - cmu-mmac.png
  - cmu-mmac.webp
  - dreamdojo.png
  - dreamdojo.webp
  - ego-exo4d.png
  - ego-exo4d.webp
  - ego4d.png
  - ego4d.webp
  - egohands.png
  - egohands.webp
  - egolife.png
  - egolife.webp
  - egoscale.png
  - egoscale.webp
  - egoschema.png
  - egoschema.webp
  - egovla.png
  - egovla.webp
  - egovlp.png
  - egovlp.webp
  - epic-kitchens-100.png
  - epic-kitchens-100.webp
  - gtea-gaze.png
  - gtea-gaze.webp
  - hoi4d.png
  - hoi4d.webp
  - hot3d.png
  - hot3d.webp
  - project-aria.png
  - project-aria.webp
  - umi.png
  - umi.webp
  - xperience-10m.png
  - xperience-10m.webp
- awesome-egocentric-access-funnel.png
- awesome-egocentric-access-funnel.svg
- awesome-egocentric-access-funnel.webp
- awesome-egocentric-atlas-cover.png
- awesome-egocentric-atlas-cover.webp
- awesome-egocentric-atlas-map.png
- awesome-egocentric-atlas-map.svg
- awesome-egocentric-atlas-map.webp
- awesome-egocentric-logo-source.png
- awesome-egocentric-logo.png
- awesome-egocentric-logo.webp
- awesome-egocentric-milestones.png
- awesome-egocentric-milestones.svg
- awesome-egocentric-reader-route.png
- awesome-egocentric-reader-route.svg
- awesome-egocentric-social-card.png
- awesome-egocentric-task-matrix.png
- awesome-egocentric-task-matrix.svg
- awesome-egocentric-timeline.png
- awesome-egocentric-timeline.svg
- awesome-egocentric-timeline.webp
- README.md
- readme_aliases.yml
- resources.yml
- taxonomy.yml
- - ar-glasses-sensing.md
  - audio-social-egocentric.md
  - hand-object-3d-hoi.md
  - long-context-egocentric-qa.md
  - robot-vla-from-ego-video.md
- access_matrix.md
- maintenance.md
- resource_schema.md
- status_policy.md
- taxonomy.md
- dataset_card_header.txt
- - catalog_artifacts.rb
- audit_catalog.rb
- build_artifacts.rb
- build_hf_package.rb
- build_readme_i18n.rb
- build_site_data.rb
- freshness.rb
- sort_readme_tables.rb
- validate_catalog.rb
- verify_hf_mirror.rb
.editorconfig
.gitignore
app.js
awesome-egocentric-atlas.csv
awesome-egocentric-papers.csv
CHANGELOG.md
CITATION.cff
CODE_OF_CONDUCT.md
CONTRIBUTING.md
feed.xml
index.html
LICENSE
README.de.md
README.es.md
README.fr.md
README.ja.md
README.ko.md
README.md
README.pt.md
README.zh.md
robots.txt
site-data.json
site.webmanifest
sitemap.xml
styles.css

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1 align="center">Awesome Egocentric Atlas</h1>

Datasets, benchmarks, models, and tools for egocentric AI.

<a href="README.md">English</a> ·
 <a href="README.zh.md">中文</a> ·
 <a href="README.es.md">Español</a> ·
 <a href="README.fr.md">Français</a> ·
 <a href="README.de.md">Deutsch</a> ·
 <a href="README.ja.md">日本語</a> ·
 <a href="README.ko.md">한국어</a> ·
 <a href="README.pt.md">Português</a> ·
 <a href="CONTRIBUTING.md">Help translate</a> ·
 <a href="https://chaoyue0307.github.io/awesome-egocentric-atlas/">Landing page</a> ·
 <a href="https://huggingface.co/datasets/cy0307/awesome-egocentric-atlas">Hugging Face mirror</a>

**Awesome Egocentric Atlas** is a practical catalog of egocentric (first-person) datasets, benchmarks, models, and tools for egocentric vision, embodied AI and robotics, vision-language-action, world models, long-context memory, AR/VR, and hand-object interaction. Every entry shows its public-access status, so you can tell at a glance what you can download today and what is still just a paper.

**Updated:** 2026-08-23.
**Scope:** the main atlas is **human or animal first-person capture** from head, glasses, headset, body, wrist, handheld, or synchronized ego-exo rigs (where the ego view is central). Related but non-egocentric resources — robot-only datasets, multi-view robotic benchmarks, autonomous-driving 4D data, and general long-video reasoning — are listed separately under [Adjacent and Related Resources](#adjacent-and-related-resources) rather than in the main tables.

## Contents

- [How to Use This Atlas](#how-to-use-this-atlas)
- [At a Glance](#at-a-glance)
- [Milestones](#milestones)
- [Start Here](#start-here)
- [Landscape Snapshot](#landscape-snapshot)
- [Creator and Release Notes](#creator-and-release-notes)
- [Status Legend](#status-legend)
- [Fast Task Map](#fast-task-map)
- [Recipes and Reference](#recipes-and-reference)
- [Dataset Atlas](#dataset-atlas)
- [Benchmarks and Derived Annotations](#benchmarks-and-derived-annotations)
- [Models, Tools, and Baselines](#models-tools-and-baselines)
- [Adjacent and Related Resources](#adjacent-and-related-resources)
- [Surveys, Papers, and Context](#surveys-papers-and-context)
- [Workshops and Challenges](#workshops-and-challenges)
- [Watchlist](#watchlist)
- [Inclusion Rules](#inclusion-rules)
- [Contributing](#contributing)
- [Cite This Atlas](#cite-this-atlas)

## How to Use This Atlas

Use the first two tables to orient yourself, then jump into the detailed atlas section that matches your task. Each entry is deliberately short: name, release date (year-month, from the first public/arXiv posting where known), venue (conference/journal where published, or `arXiv`/source for preprints), scale/signal, best use, and public status. For filtering by task, modality, status, or date, use the catalog in [`data/resources.yml`](data/resources.yml), where every entry carries a `released` field; label meanings and filter groups are explained in [docs/taxonomy.md](docs/taxonomy.md). Task-oriented [research recipes](#recipes-and-reference) walk you from goal to experiment.

Prefer a browsable view? The [interactive site](https://chaoyue0307.github.io/awesome-egocentric-atlas/) lets you filter the catalog by task, status, and date in eight languages, and the same catalog is mirrored as a dataset on [Hugging Face](https://huggingface.co/datasets/cy0307/awesome-egocentric-atlas).

## At a Glance

| Signal | What it means for readers |
| :--- | :--- |
| 866 egocentric resources | 229 datasets, 142 benchmarks, 442 models, and 43 toolkits, plus 4 collection hubs — across vision, robotics, memory, and AR. 178 related non-egocentric resources are listed separately. |
| 6 research areas | Foundation video, procedure/action, hands and 3D, memory/reasoning, robotics/VLA, and AR/wearable sensing. |
| 5 access states | `open`, `request`, `benchmark`, `partial`, and `watch` keep availability visible before you plan experiments. |
| Machine-checked catalog | [`data/resources.yml`](data/resources.yml) is the source for type, year, status, URL, tasks, and provenance — and CI keeps the public artifacts in sync. |
| Reader-first tables | Each entry is short enough to scan, then links out to the official page, paper, code, or dataset portal. |

## Milestones

Representative works to read first, from early egocentric activity datasets to recent models and large-scale corpora.

## Start Here

| Reader | Best first move | Then go deeper |
| :--- | :--- | :--- |
| New to egocentric vision | Read Ego4D, EPIC-KITCHENS-100, Ego-Exo4D, and EgoSchema first. | Add EgoClip/EgoVLP for video-language, VISOR for masks, and HOT3D/HOI4D for 3D hand-object work. |
| Building a benchmark table | Start from the `open` and `benchmark` entries, then inspect `request` licenses. | Use the YAML catalog to filter by task and modality. |
| Building an assistant or agent | Start with Ego4D, Xperience-10M, EgoLife, TeleEgo, EgoSchema, MyEgo, EgoIntrospect, EgoBench, UCS-Bench, and Ego2Web. | Track Ego-R1, X-LeBench, EgoMemReason, StreamMemBench, ReFocus, EgoCoT-Bench, EgoBlind, EgoSound, and NoRA for reasoning diagnostics. |
| Building a robot/VLA system | Start with Xperience-10M, EgoDex, EgoVerse, Ego-Exo4D, HoloAssist, HOI4D, HOT3D, Assembly101, OpenEgo, UMI, FastUMI-100K, EgoPAT3D, HUG, and WiYH. | Add Dobb-E, FastUMI, MV-UMI, UMIGen, YUBI, Hoi!, EgoVLA, EgoEngine, HumanEgo, EgoGuide, Ego-Pi, EgoAERO, H2O, ARCTIC, HOMIE-toolkit, Xperience task baselines, and EgoHandTrajPred for policy/action supervision. For robot-only corpora see [Adjacent and Related](#adjacent-and-related-resources). |
| Studying hands, objects, and 3D | Start with HOT3D, HOI4D, H2O, ARCTIC, FPHA, EgoHOS, EgoBody, GIMO, EgoHumans, EgoTactile, and UnrealEgo. | Add EgoTracks/TREK-150 for tracking, EgoForce for monocular 3D hand pose, and EgoEMG/EgoEVHands/EgoPressDiff for emerging sensors and contact. |

## Landscape Snapshot

> **Origins.** Egocentric vision grew out of wearable computing: [Steve Mann](http://wearcam.org/) pioneered continuous first-person ("sousveillance") cameras through the 1980s–90s and is widely credited with opening up the field, and Microsoft's SenseCam (2006) popularized lifelogging capture. The modern computer-vision wave began at the first IEEE Workshop on Egocentric Vision (CVPR 2009), where Spriggs, De La Torre, and Hebert introduced egocentric activity recognition on the [CMU-MMAC](http://kitchen.cs.cmu.edu/) database (2009) — the earliest dataset in this atlas — followed by the foundational GTEA (2011) and ADL (2012) datasets that launched first-person activity and object research.

## Creator and Release Notes

| Resource | Created / stewarded by | Release channel and access note |
| :--- | :--- | :--- |
| [Xperience-10M](https://huggingface.co/datasets/ropedia-ai/xperience-10m) | Ropedia | Released under `ropedia-ai` on Hugging Face. Full dataset uses controlled non-commercial access; the [sample](https://huggingface.co/datasets/ropedia-ai/xperience-10m-sample) is open, [HOMIE-toolkit](https://github.com/Ropedia/HOMIE-toolkit) loads/visualizes annotations, and [Xperience task baselines](https://huggingface.co/cy0307/ropedia-xperience-10m-task-baselines) provide a public evaluation artifact. |
| [Ego4D](https://ego4d-data.org/) | Meta AI / Facebook AI with an international academic consortium | Official data portal with license-gated downloads and benchmark tasks. |
| [Ego-Exo4D](https://ego-exo4d-data.org/) | Meta and academic partners across the Ego-Exo4D consortium | Official data portal with application/download tooling for synchronized ego-exo data. |
| [EPIC-KITCHENS](https://epic-kitchens.github.io/) | University of Bristol-centered EPIC-KITCHENS team | Official challenge/dataset portal with public annotations and benchmark splits. |
| [Project Aria datasets](https://www.projectaria.com/datasets/) | Meta / Project Aria | Official Project Aria dataset portal and tooling for AR-glasses sensor data. |

## Status Legend

For how each status is decided, see [docs/status_policy.md](docs/status_policy.md).

## Fast Task Map

| Research goal | Strong starting points |
| :--- | :--- |
| General egocentric foundation data | [Xperience-10M](https://huggingface.co/datasets/ropedia-ai/xperience-10m), [Ego4D](https://ego4d-data.org/), [Ego-Exo4D](https://ego-exo4d-data.org/), [EPIC-KITCHENS-100](https://epic-kitchens.github.io/), [EgoClip](https://github.com/showlab/EgoVLP) |
| Ego-exo and skill learning | [Ego-Exo4D](https://ego-exo4d-data.org/), [EgoExoLearn](https://github.com/OpenGVLab/EgoExoLearn), [Assembly101](https://assembly-101.github.io/), [Look and Tell](https://arxiv.org/abs/2510.22672) |
| Robot manipulation / VLA | [Xperience-10M](https://huggingface.co/datasets/ropedia-ai/xperience-10m), [EgoDex](https://arxiv.org/abs/2505.11709), [EgoVerse](https://egoverse.ai/), [OpenEgo](https://arxiv.org/abs/2509.05513), [UMI](https://umi-gripper.github.io/), [FastUMI-100K](https://huggingface.co/datasets/IPEC-COMMUNITY/FastUMI_100k_lerobot), [YUBI](https://arxiv.org/abs/2606.10244), [HUG](https://grasping.io/), [EgoVLA](https://rchalyang.github.io/EgoVLA/), [EgoEngine](https://egoengine.github.io/), [HumanEgo](https://humanego-ai.github.io) |
| Long-context QA and memory | [EgoSchema](http://egoschema.github.io/), [EgoLife](https://arxiv.org/abs/2503.03803), [EgoMemReason](https://arxiv.org/abs/2605.09874), [UCS-Bench](https://github.com/cocowy1/UCS-Bench), [StreamMemBench](https://arxiv.org/abs/2606.14571), [Ego-EXTRA](https://fpv-iplab.github.io/Ego-EXTRA/), [EgoStream](https://arxiv.org/abs/2605.31557), [EgoBench](https://arxiv.org/abs/2605.27820) |
| Hand-object and 3D HOI | [HOT3D](https://facebookresearch.github.io/hot3d/), [HOI4D](https://hoi4d.github.io/), [EgoHOS](https://github.com/owenzlz/EgoHOS), [FPHA](https://guiggh.github.io/publications/first-person-hands/), [HUG](https://grasping.io/), [EgoTactile](https://egotactile.github.io/), [EgoForce](https://dfki-av.github.io/EgoForce), [EgoEMG](https://arxiv.org/abs/2605.05712) |
| Audio, speech, and social interaction | [Ego4D AV tasks](https://ego4d-data.org/), [EasyCom](https://arxiv.org/abs/2107.04174), [EgoCom / audio-visual correspondence](http://vision.cs.utexas.edu/projects/ego_av_corr/), [AV-CONV](https://vjwq.github.io/AV-CONV/) |
| AR glasses and scene understanding | [Project Aria Datasets](https://www.projectaria.com/datasets/), [Aria Digital Twin](https://www.projectaria.com/datasets/adt/), [AEA](https://www.projectaria.com/datasets/aea/), [Nymeria](https://www.projectaria.com/datasets/nymeria/), [DTC](https://www.projectaria.com/datasets/dtc/) |

## Recipes and Reference

Task-oriented starting points and the references that keep the catalog usable.

**Research recipes** — each gives a goal, what to start with, what to add, benchmarks, baselines, a minimum viable experiment, common pitfalls, and a reporting checklist:

- [Long-context egocentric video QA](docs/recipes/long-context-egocentric-qa.md)
- [Robot / VLA from egocentric human video](docs/recipes/robot-vla-from-ego-video.md)
- [Hand-object 3D HOI](docs/recipes/hand-object-3d-hoi.md)
- [AR-glasses and wearable sensing](docs/recipes/ar-glasses-sensing.md)
- [Audio and social egocentric understanding](docs/recipes/audio-social-egocentric.md)

**Reference**

- [Taxonomy](docs/taxonomy.md) — task/modality families and how to filter the catalog.
- [Access Matrix](docs/access_matrix.md) — data, code, license, and leaderboard per key resource.
- [Status Policy](docs/status_policy.md) — how `open`/`request`/`benchmark`/`partial`/`watch` are decided.
- [Resource Schema](docs/resource_schema.md) — every field in `data/resources.yml`.
- [Maintenance](docs/maintenance.md) — the add → verify → validate workflow.

## Dataset Atlas

### Foundational and Large-Scale Ego Video

Broad, large-scale first-person corpora that most egocentric pipelines pretrain or benchmark on.

| Resource | Released | Venue | Scale / signal | Best for | Status |
| :--- | :---: | :---: | :--- | :--- | :---: |
| [RekaDaily-10k (raw)](https://huggingface.co/datasets/RekaAI/RekaDaily-10k-raw) | 2026-08 | Hugging Face | Current incremental release: 7,834 hours, 397,171 unscripted first-person videos, 9,836 WebDataset shards, and about 70 TB | Large-scale daily-life pretraining, action understanding, and human-to-robot learning | open |
| [Ego500](https://huggingface.co/datasets/humanarchive/ego500) | 2026-08 | Hugging Face | Announced 500-hour corpus with 310K+ structured action annotations across 390 task-environment combinations and 60+ environments; manual access and upload are still pending | Dense work-procedure, hand-state, and task-phase modeling | request |
| [EgoSuite-Open100K](https://huggingface.co/collections/LightwheelAI/egosuite-open100k) | 2026-08 | Hugging Face | Gated EgoDemo, EgoStandard, and EgoPro buckets advertise 100K hours across head and head-plus-wrist lines with hand/body pose and semantics; 100K remains the collection target, while current manifests are the release truth | Large-scale controlled-access egocentric pretraining and VLA data | request |
| [Stereo-550](https://huggingface.co/datasets/fpvlabs/stereo-550) | 2026-07 | Hugging Face | 550 hours per camera, 1,462 calibrated stereo sessions, 209,315 dense action segments, and synchronized 6-axis IMU for 1,271 sessions | Calibrated stereo HOI, dense video-language learning, and sensor fusion | open |
| [ACE-Data-0](https://ace-data-engine.github.io/ACE-Data-0/) | 2026-07 | arXiv | Current public snapshot has 30 room-scale takes, 95,238 files, and about 3.61 TB of ego/exo video and multimodal sidecars; 150 hours and 75K+ interactions remain the full-release target | Multisensory household imitation, world models, and VLA research | partial |
| [RetailSMV](https://arxiv.org/abs/2607.00310) | 2026-06 | arXiv | 32,105 captioned synchronized retail clips from five supermarkets with staff-perspective egocentric and exocentric views | Retail world-model adaptation and ego/exo viewpoint comparison | watch |
| [EgoCS-400K](https://arxiv.org/abs/2606.18180) | 2026-06 | arXiv | 400K+ first-person Counter-Strike gameplay videos (10K hours, 13 maps) with aligned actions, player state, camera motion, and game events | Action-conditioned interactive world models from first-person gameplay | watch |
| [HumanNet](https://arxiv.org/abs/2605.06747) | 2026-05 | arXiv | 1M-hour human-centric video corpus spanning first-person and third-person views, with captions, motion descriptions, hand/body signals, and a reported 1K-hour egocentric subset for VLA-style validation | Large-scale human-video pretraining, interaction understanding, and human-to-robot transfer | watch |
| [EgoInteract](https://huggingface.co/datasets/EgoInteract/EgoInteract) | 2026-05 | arXiv | Public synthetic release with 10,534 episodes, about 1.9M frames, 61,211 NAO/HOI task frames, and released simulator/training code | Synthetic temporal segmentation, next-active-object detection, anticipation, and EHOI | open |
| [PRISM](https://dreamvu.ai/prism) | 2026-03 | arXiv | Paper reports 270K multi-view retail SFT samples; the gated PRISM-100K release provides 100K samples, 25,173 ego/exo videos, 21 task types, and about 772 GB | Embodied VLM spatial, physical, and action reasoning in retail settings | request |
| [Ego-1K](https://huggingface.co/datasets/facebook/ego-1k) | 2026-03 | CVPR 2026 | Nearly 1,000 synchronized multiview egocentric videos from a custom 12-camera plus VR-headset rig | Dynamic 3D/4D scene understanding and novel view synthesis from ego rigs | open |
| [EgoCrowds / CrowdEraser](https://arxiv.org/abs/2603.29036) | 2026-03 | arXiv | Semi-synthetic paired crowded/empty clips from real egocentric walking-tour video; CrowdEraser diffusion removes crowds for humanless walkthroughs | First-person walking-tour video editing and environment modeling | watch |
| [Xperience-10M](https://huggingface.co/datasets/ropedia-ai/xperience-10m) | 2026-03 | Hugging Face | Ropedia release on Hugging Face; 10M experiences, 10K hours, six video streams, audio, stereo depth, camera pose, hand/body mocap, IMU, hierarchical language, ~1 PB total | Embodied AI, world models, robot learning from human experience, sensor fusion, 3D/4D understanding | request |
| [Xperience-10M Sample](https://huggingface.co/datasets/ropedia-ai/xperience-10m-sample) | 2026-03 | Hugging Face | Public sample episode for Xperience-10M with six rows on Hugging Face and `cc-by-nc-4.0` terms | Loader testing, demos, task-suite prototyping, annotation inspection | open |
| [Egocentric-100K](https://huggingface.co/datasets/builddotai/Egocentric-100K) | 2025-12 | Hugging Face | 100,405-hour manual-labor egocentric video corpus with 10.8B frames, 2.01M clips, intrinsics, and WebDataset shards | Industrial/manual-labor pretraining, robot learning, and active manipulation density analysis | request |
| [Egocentric-10K](https://huggingface.co/datasets/builddotai/Egocentric-10K) | 2025-11 | Hugging Face | 10,000-hour factory-only egocentric video corpus with 1.08B frames, intrinsics, and WebDataset-style shards | Industrial egocentric pretraining and active-manipulation density analysis | request |
| [Voxel51 Egocentric-10K Subset](https://huggingface.co/datasets/Voxel51/Egocentric_10K_subset) | 2025-11 | Hugging Face | Open FiftyOne package for a Factory 51 subset of Egocentric-10K, with the card reporting 416 1080p head-mounted industrial clips | Lightweight inspection, prototyping, and visualization of Egocentric-10K-style factory footage | open |
| [Look and Tell](https://arxiv.org/abs/2510.22672) | 2025-10 | NeurIPS 2025 Workshop | 25 participants, Project Aria plus stationary cameras, gaze/speech/video, 3D reconstructions | Referential communication across ego and exo viewpoints | watch |
| [EgoBlind](https://arxiv.org/abs/2503.08221) | 2025-03 | NeurIPS 2025 D&B | 1,392 first-person videos from blind and visually impaired users, 5,311 questions posed or verified by blind users | Assistive egocentric VideoQA for blind users | watch |
| [HD-EPIC](https://arxiv.org/abs/2502.04144) | 2025-02 | CVPR 2025 | 41 hours, 9 kitchens, dense fine-grained labels, 3D fixture annotations, audio events, VQA | Detailed kitchen understanding, VQA, 3D-aware ego reasoning, audio-event recognition | open |
| [EgoVid-5M](https://egovid.github.io/) | 2024-11 | NeurIPS 2025 | 5M curated egocentric clips at 1080p with fine-grained kinematic and high-level text action annotations (NeurIPS 2025) | Egocentric video generation and action-conditioned world models | open |
| [Ego-Exo4D](https://ego-exo4d-data.org/) | 2023-11 | CVPR 2024 | 1,286 hours, 740 participants, synchronized first- and third-person views, audio, gaze, IMU, 3D point clouds, language commentary | Skilled activity, ego-exo transfer, cross-view translation, pose, proficiency, robotics | request |
| [HoloAssist](https://holoassist.github.io/) | 2023-09 | ICCV 2023 | 166 hours, 350 instructor-performer pairs, 7 synchronized streams from mixed-reality headsets | Interactive assistance, mistake detection, intervention prediction, procedural collaboration | open |
| [EgoObjects](https://github.com/facebookresearch/EgoObjects) | 2023-09 | ICCV 2023 | Pilot release with 9K+ videos, 250 participants, 50+ countries, 650K annotations, 368 categories | Category and instance-level egocentric object understanding, continual object detection | open |
| [EgoCom / Ego audio-visual correspondence](http://vision.cs.utexas.edu/projects/ego_av_corr/) | 2023-07 | CVPR 2024 | Egocentric video with spatial audio for conversation and audio-visual correspondence tasks | Active speaker detection, spatial audio denoising, conversational graph reasoning | open |
| [Assembly101](https://assembly-101.github.io/) | 2022-03 | CVPR 2022 | 4,321 videos, 8 static plus 4 egocentric views, 100K coarse and 1M fine action segments, 18M 3D hand poses | Procedural action, anticipation, mistake detection, cross-view transfer | open |
| [Ego4D](https://ego4d-data.org/) | 2021-10 | CVPR 2022 | 3,670+ hours from 900+ camera wearers, multiple countries, video/audio/gaze/stereo/3D/narrations depending on subset | Long-form ego video, episodic memory, social, hand-object, forecasting, audio-visual tasks | request |
| [EasyCom](https://arxiv.org/abs/2107.04174) | 2021-07 | arXiv | AR glasses egocentric multi-channel audio and wide-FOV RGB for noisy conversations | Speech enhancement, source localization, conversation assistance | open |
| [EPIC-KITCHENS-100](https://epic-kitchens.github.io/) | 2020-06 | IJCV 2022 | 100 hours, 20M frames, 90K actions, 45 kitchens, narrations and dense action labels | Kitchen action recognition, anticipation, action detection, retrieval | open |
| [From Third Person to First Person](https://arxiv.org/abs/1812.00104) | 2018-12 | CVPR 2019 | Ego/exo video synthesis and retrieval datasets plus baselines for bridging third-person social video to first-person views | Cross-view synthesis, retrieval, and ego-exo transfer | watch |
| [EPIC-KITCHENS](https://epic-kitchens.github.io/) | 2018-04 | ECCV 2018 | 55 hours / 11.5M frames from 32 kitchens with narrations, 39.6K action segments, and object boxes | Original large-scale kitchen action recognition and anticipation benchmark | open |

### Robotics, Manipulation, and VLA

Egocentric human-manipulation and wrist-camera data aimed at vision-language-action models and human-to-robot transfer.

| Resource | Released | Venue | Scale / signal | Best for | Status |
| :--- | :---: | :---: | :--- | :--- | :---: |
| [Egocentric Manufacturing](https://huggingface.co/datasets/Worlddatalabs/egocentric-manufacturing) | 2026-08 | Hugging Face | 259 gated clips, about 100 hours, 14 live-factory process domains, 256 workers, 3,911 scenes, and 2,081 entities | Industrial procedures, VLA pretraining, and authentic plant-floor HOI | request |
| [Axis Ego Samples](https://huggingface.co/datasets/axisrobotics/axis-ego-samples) | 2026-08 | Hugging Face | Public 100-hour LeRobot egocentric sample across nine-plus work domains, plus ten episodes from a separate 150-hour delivery corpus; reuse terms are unstated | Broad workplace human-demonstration and loader experiments | open |
| [RealFactory-Ego-Data](https://huggingface.co/datasets/FlyBirdData/RealFactory-Ego-Data) | 2026-08 | Hugging Face | Five Apache-2.0 head-mounted episodes totaling about 14 minutes on a real touchscreen line, with RGB, 21-point hands, and frame-level actions | Compact industrial HOI and action-segmentation prototypes | open |
| [BaseMatrix EGO Desktop Demo](https://huggingface.co/datasets/basematrix/ego-desktop-demo) | 2026-08 | Hugging Face | One 70-second, 2,114-frame binocular ego/exo episode with camera pose, metric 3D hands, contact, and derived actions | Inspecting a dense stereo hand-retargeting schema | open |
| [ActTrace Ego-Home](https://huggingface.co/datasets/ActTrace/acttrace-ego-home) | 2026-08 | Hugging Face | 13 household episodes over 11 tasks and 18 minutes with RGB, LiDAR depth, 100 Hz IMU, metric hands, and IMU-verified 6-DoF head trajectories | Metric RGB-D/IMU household demonstrations and sensor fusion | open |
| [Everyday Manipulation 3D](https://caryx.ai/data) | 2026-08 | Hugging Face | 116 densely annotated episodes and 28,019 frames with depth, pose, MANO, contact, masks, and language; companion raw release has 1,513 clips, 10.28 hours, and 279 GiB | Dense 3D HOI evaluation plus raw RGB-D pretraining | open |
| [VITRA Egocentric Instruction Sidecars](https://huggingface.co/datasets/MIT-Media-Lab/vitra-instructions-qwen35-122b-egocentric-v1) | 2026-08 | Hugging Face | 95,831 accepted hand-action instructions from 175,327 Qwen3.5 candidates across GigaHands, HOT3D, and OakInk2; source revisions and paths are preserved, but no standalone license is declared | VLA instruction tuning and annotation-sidecar research | partial |
| [InHandPlus Wrist-View Camera and IMU Sample](https://huggingface.co/datasets/inhandplus/egocentric-wrist-view-camera-IMU-data) | 2026-08 | Hugging Face | Auto-gated 30-episode cloth-manipulation sample with wrist video, synchronized raw 6-axis IMU, and LeRobot v3 metadata under CC-BY-NC-4.0 | Wrist-view sensor fusion and early loader experiments | request |
| [InfoBayAI Egocentric Video Sample](https://huggingface.co/datasets/InfoBayAI/egocentric_video) | 2026-08 | Hugging Face | 10 gated MP4 samples and metadata totaling 31.87 GB; the much larger headline collection on the card is not present in the repository manifest | Inspecting first-person sample media without treating the advertised corpus as released | request |
| [UniDataPro Egocentric Video Sample](https://huggingface.co/datasets/UniDataPro/egocentric-video) | 2026-08 | Hugging Face | One public 2.33 GB video/metadata/tracking preview for a commercial collection advertised as 4,050 hours of Pico, ZED, IMU, and motion-tracker capture | Evaluating a commercial multimodal manipulation-data offering | request |
| [Nexdata 10,000-Hour Egocentric Video Dataset](https://huggingface.co/datasets/Nexdata-AI/10000-Hour-Egocentric-Video-Dataset) | 2026-08 | Hugging Face | Commercial specification for 10,000 hours of 4K stereo head video, wrist/ankle IMU, 76-point body pose, and dense actions; full media is available by request | Evaluating a large commercial multimodal capture offering | request |
| [ROCO IROS 2026 UMI Dataset](https://huggingface.co/datasets/rocochallenge2025/roco_iros2026_umi_dataset) | 2026-08 | Hugging Face | 4,928 public LeRobot episodes and 7.46M frames with stereo FPV, four hand cameras, and bimanual pose/gripper state; no card or license is supplied | Bimanual UMI/VLA loading and six-camera policy experiments | partial |
| [DreamTraj / MOVE](https://whathappen0.github.io/DreamTraj/) | 2026-08 | arXiv | 5,038 language-conditioned egocentric 6-DoF object trajectories; predicts motion from unrendered video-diffusion latents at 4.6x the reported speed of generate-then-extract pipelines | Object-trajectory prediction and language-to-manipulation motion | watch |
| [Ego2Robot](https://www-ye.github.io/ego2robot_blog/) | 2026-08 | arXiv | Converts egocentric human manipulation video into 18,561 hours of synthesized robot-format data across 15 morphologies | Scalable ego-to-robot pretraining and embodiment generalization | watch |
| [RoboReact](https://arxiv.org/abs/2608.03387) | 2026-08 | arXiv | Generates egocentric RGB-D manipulation plans, reconstructs interaction keyframes, retargets them to humanoids, and refines execution in closed loop | Generated-video-to-humanoid whole-body skill distillation | watch |
| [SiMDex](https://lin-nie.github.io/SiMDex/) | 2026-08 | arXiv | Mines 1.49M task-relevant samples from roughly 32M egocentric clips and reports 47.7% to 61.1% dexterous success over equal-sized random sampling | Relevance-aware human-video curation for dexterous VLA post-training | watch |
| [VLAff / EgoAffordance](https://ojh6404.github.io/vlaff/) | 2026-08 | IROS 2026 | 204K episodes with 5.6M visual and 11.6M grasp/trajectory affordances; predicts heatmaps, grasps, and executable trajectories | Actionable affordance learning and human-to-robot transfer | watch |
| [JoyAI-RA 0.5](https://joyai-ra-05.github.io/) | 2026-08 | arXiv | Dual latent/canonical action alignment across human egocentric video, simulation, and robot data for a generalist VLWA policy | Scaling manipulation learning from heterogeneous human and robot experience | watch |
| [HumynLabs Egocentric Sample Collection](https://huggingface.co/humyn-labs) | 2026-08 | Hugging Face | Five CC-BY-4.0 packs covering head/dual-wrist video and IMU, stereo/mono labels, voice-over captions, and regional residential IMU | Small multimodal capture, sensor-fusion, and loader experiments | open |
| [EgoGenesis](https://egogenesis.github.io/) | 2026-07 | arXiv | Geometry-aware egocentric world-action simulator; 400 generated plus 400 real trajectories improve reported OOD success from 77% to 84% single-arm and 53% to 70% dual-arm | Controllable ego-video generation and world-action-model augmentation | watch |
| [Firstly Household Manipulation v0.1](https://huggingface.co/datasets/FirstlyLabs/Firstly-household-v01) | 2026-07 | Hugging Face | 12 open dishwashing/laundry episodes totaling 119.1 minutes and 214,437 frames, with 148 phase instructions and two-hand tracks | Compact LeRobot human-demonstration and annotation-pipeline testing | open |
| [HiFi-UMI-2K](https://huggingface.co/datasets/simple-world-lab/HiFi-UMI-2K) | 2026-07 | arXiv | 2,000 open hours from a high-fidelity UMI rig with head stereo-inertial SLAM, native inter-gripper pose, microsecond synchronization, and two ultra-wide cameras per hand | Robot-free VLA/WAM pretraining and direct manipulation-policy deployment | open |
| [EgoRecovery](https://arxiv.org/abs/2607.19745) | 2026-07 | arXiv | Egocentric human recovery demonstrations collected at more than 10x robot-teleoperation throughput and aligned with a shared corrective-intent space | Learning when and how robots should recover from manipulation failures | watch |
| [DDD Egocentric Stereo Manipulation Sample](https://huggingface.co/datasets/DDD-Cambodia/ego-centric-sample-dataset) | 2026-07 | Hugging Face | Ten open LeRobot v2 episodes (about 2.37 GB) with stereo first-person video and high-frequency hand/head pose | Stereo human-demonstration loaders, hand tracking, and sensor fusion | open |
| [Showway Egocentric Origami Series](https://huggingface.co/datasets/showway-ego/egocentric-origami-001) | 2026-07 | Hugging Face | Nine open neck-mounted iPhone origami episodes across three LeRobot v3 repositories, with 3D hand joints and reviewed narration | Compact procedural HOI and imitation-learning experiments | open |
| [EGXO Household Egocentric Video Evaluation](https://huggingface.co/datasets/egxodata/egxo-household-egocentric-video-evaluation) | 2026-07 | Hugging Face | 71 household-task videos totaling 10 hours; six previews are public and the complete 43.4 GiB media package is licensed by request | Household procedure and hand-object video evaluation | request |
| [Let the Body Follow](https://arxiv.org/abs/2607.16095) | 2026-07 | arXiv | Coupled egocentric TIAGo teleoperation maps head and arm motion to coordinated torso and mobile-base control, reducing explicit controls and user workload | Whole-body mobile-manipulator teleoperation and HRI | watch |
| [EgoHTR](https://egohtr.github.io/) | 2026-07 | arXiv | 55 scene-aligned 4D human-terrain sequences across seven environments, totaling 1.37 hours and about 150K frames from eight participants, with multiview and mocap ground truth | Scene-aware human-motion reconstruction and humanoid terrain traversal | watch |
| [Open-AoE](https://huggingface.co/datasets/inclusionAI/OpenAoE-2000h) | 2026-07 | Hugging Face | Nano 3-hour and tiny 100-hour subsets are released; about 323 hours were uploaded and 694 hours prepared by August 12, with 11.9 TB currently stored; 2,000 hours remains the roadmap | Large-scale hand-motion, camera-trajectory, and manipulation pretraining | partial |
| [SenseXperience Raw MCAP Sample](https://huggingface.co/datasets/io-intelligence/SenseXperience_WristCam_SampleData) | 2026-07 | Hugging Face | 12 public raw ROS 2 MCAP episodes with four synchronized cameras, head IMU near 120 Hz, calibration, metadata, and recorder diagnostics | Raw multimodal capture-pipeline and sensor-synchronization testing | open |
| [Ego Data by Object](https://huggingface.co/datasets/Kavin60606/ego-data-by-object) | 2026-07 | Hugging Face | 4,110 pick-and-place episodes over 51 object labels with egocentric RGB and HDF5 hand/body transforms, intrinsics, task attributes, and descriptions; source licensing is unclear | Object-conditioned hand tracking and imitation-data experiments | partial |
| [EgoExo-Wreck](https://huggingface.co/datasets/WreckVegas/EgoExo-Wreck) | 2026-07 | Hugging Face | Commercial card reports 150+ synchronized ego-exo hours of object destruction, debris handling, cleanup, and reset; only a gated 1 GB evaluation sample is hosted | Physics-rich HOI, state-change, world-model, and cleanup demonstrations | request |
| [Verbose Industrial Egocentric Sample Series](https://huggingface.co/VerboseTechLabs) | 2026-07 | Hugging Face | Six open industrial sample repositories with 53 first-person clips, 4.86 hours, and about 17.0 GB across textile, metal, electronics assembly, skilled work, and manufacturing operations | Compact industrial action and human-demonstration experiments | open |
| [Verbose Cooking and Chopping Egocentric Video](https://huggingface.co/datasets/VerboseTechLabs/cooking-chopping-egocentric) | 2026-07 | Hugging Face | 11 head- or chest-mounted kitchen clips totaling 101.7 minutes in a public 6.09 GB ZIP, with audio and clip metadata across cooking and chopping | Open kitchen action, procedure, and hand-object video experiments | open |
| [Verbose Household Cleaning Egocentric Video](https://huggingface.co/datasets/VerboseTechLabs/household-cleaning-egocentric) | 2026-07 | Hugging Face | 12 head- or chest-mounted 1080p/30 fps cleaning clips totaling 104.3 minutes in a public 5.74 GB ZIP, with audio and three sub-activity labels | Open domestic-action and assistive-robotics video experiments | open |
| [Flikforge Egocentric Household Tasks](https://huggingface.co/datasets/FlikforgeInc/egocentric-household-tasks) | 2026-07 | Hugging Face | Card claims about 100 rights-cleared household clips with dense customizable annotations, but the repository currently has no media or annotation files and its LICENSE is empty | Monitoring a prospective commercial household-demonstration sample release | watch |
| [AgenticFocus](https://arxiv.org/abs/2607.08857) | 2026-07 | arXiv | Restores occluded object geometry and full-hand motion from ordinary first-person human videos, then retargets and composites robot-aligned observations, actions, and states | Mixed-reality human-to-humanoid demonstration synthesis | watch |
| [Hub Egocentric GoPro RGB+IMU](https://huggingface.co/datasets/Hubdata/egocentric-gopro-rgb-imu) | 2026-07 | Hugging Face | 19 GoPro HERO13 manipulation clips totaling 3.72 hours, with RGB, approximately 200 Hz GPMF IMU, sensor sidecars, and MCAP | High-rate visual-inertial human-demonstration prototyping | partial |
| [Hub Egocentric iPhone RGB+IMU](https://huggingface.co/datasets/Hubdata/egocentric-iphone-rgb-imu) | 2026-07 | Hugging Face | 12 iPhone manipulation clips totaling 2.96 hours, with 1080p RGB, CoreMotion/ARKit IMU and attitude, intrinsics, sidecars, and MCAP | Smartphone RGB-IMU synchronization and sensor-fusion experiments | partial |
| [Hub Egocentric Stereo RGB-D](https://huggingface.co/datasets/Hubdata/egocentric-stereo-rgbd) | 2026-07 | Hugging Face | 9 ZED X Mini clips with stereo RGB, metric depth, IMU, 6DoF VIO, calibration, MCAP, and per-clip LeRobot v3.0 packages | RGB-D/VIO pipelines and multimodal imitation-learning loaders | partial |
| [Hub Egocentric iPhone RGB](https://huggingface.co/datasets/Hubdata/egocentric-iphone-rgb) | 2026-07 | Hugging Face | 40 household-manipulation clips totaling 1.60 hours, delivered as 1080p H.264 RGB without IMU, depth, pose, or robot actions | Lightweight first-person manipulation-video baselines | partial |
| [Datoric Residential Egocentric Video](https://huggingface.co/datasets/Datoric/egocentric-residential-video-100000h) | 2026-07 | Hugging Face | Commercial card claims 100,000 hours of 1080p+ household first-person video from Brazilian homes with task, action, object, HOI, and completion-state annotations; public files are specification/sample metadata | Reviewing large-scale household capture specifications and requesting production data | request |
| [Datoric Industrial Egocentric Video](https://huggingface.co/datasets/Datoric/industrial-egocentric-video-50000h) | 2026-07 | Hugging Face | Commercial card claims 50,000 hours of workplace first-person video across logistics, inspection, assembly, and cleaning; public files are specification/sample metadata | Reviewing industrial capture specifications and requesting production data | request |
| [EgoWAM](https://gatech-rl2.github.io/egowam.github.io/) | 2026-07 | arXiv | Controlled human-robot WAM co-training study comparing pixel, DINO, and 3D-flow targets on three real bimanual tasks, with up to 4x OOD gains reported for DINO | Human-to-robot world-representation transfer | watch |
| [USA Egocentric](https://huggingface.co/datasets/Grably/usa-egocentric) | 2026-07 | Hugging Face | Three iPhone kitchen-manipulation episodes with RGB videos, camera trajectories, hand outputs, action segments, object/contact tracks, and quality reports | Open manipulation-pipeline debugging and VLA annotation prototyping | open |
| [RynnWorld-4D](https://arxiv.org/abs/2607.06559) | 2026-07 | arXiv | 4D embodied world model over future RGB, depth, and optical flow; reports Rynn4DDataset 1.0 with 254.4M+ frames across egocentric human and robotic manipulation videos | Robot manipulation world modeling and action-conditioned 4D prediction | watch |
| [RynnWorld-Teleop](https://arxiv.org/abs/2607.06558) | 2026-07 | arXiv | Digital-teleoperation world model driven by operator hand-pose streams to synthesize egocentric videos and action labels | Scalable bimanual robot data generation from human pose streams | watch |
| [LingBot-VLA 2.0](https://arxiv.org/abs/2607.06403) | 2026-07 | arXiv | VLA update trained with around 60K hours, including 10K hours of egocentric human videos and 50K hours of robot trajectories | Whole-body VLA pretraining and practical robot deployment | watch |
| [LingBot-Video](https://arxiv.org/abs/2607.07675) | 2026-07 | arXiv | MoE DiT video foundation model for embodied intelligence, adding robot-oriented manipulation/navigation and egocentric-perspective footage plus physically grounded rewards | Embodied video pretraining and robot/world-model generation | watch |
| [LIME](https://arxiv.org/abs/2607.02417) | 2026-07 | arXiv | Intent-aware camera-motion generator mined from egocentric video, pairing language intents with relative SE(3) target poses | Active camera control and next-view prediction for embodied agents | watch |
| [H-Tac / TTP](https://arxiv.org/abs/2607.01067) | 2026-07 | arXiv | 160-hour egocentric human tactile-action dataset with 300+ tasks and 135K episodes, plus transferable tactile pretraining | Contact-rich dexterous manipulation from human tactile demonstrations | watch |
| [HomER v2](https://huggingface.co/datasets/toloka/homer-v2) | 2026-06 | Hugging Face | 765 public first-person household videos totaling 99.60 hours from 420 participants across 49 countries and 10 activity categories | Diverse everyday activity, hand-object interaction, video-language, and robot-learning research | open |
| [Ego-Tactile Manipulation](https://huggingface.co/datasets/OpenGraphLabs-Research/ego-tactile-manipulation) | 2026-06 | Hugging Face | 72 public episodes totaling 1.28 hours and 138,629 frames with 1080p ego video, 160-channel dual-hand pressure, head/wrist IMU, and contact-derived action segments | Open tactile-grounded manipulation, contact, and action-segmentation research | open |
| [EgoSteer](https://egosteer.github.io/) | 2026-06 | arXiv | Open full-stack dexterous VLA system using 9.6K hours of curated egocentric human video, with EgoSmith, a robot stack, and two released 3B checkpoints | Steerable dexterous manipulation and human-video-to-robot pretraining | open |
| [Poseidon First-Person Task Video Samples](https://huggingface.co/datasets/psdn-ai/first-person-task-video-samples) | 2026-06 | Hugging Face | Five auto-gated 1080p clips totaling 174 seconds across kitchen, gardening, maintenance, fine-motor, and cleaning tasks | Reviewing task diversity, hand-object visibility, and capture quality | request |
| [WT-UMI](https://arxiv.org/abs/2606.13232) | 2026-06 | arXiv | Wearable whole-body tactile UMI interface aligning human demonstrations and teleoperation through tactile images, contact forces, and end-effector poses | Contact-aware humanoid whole-body manipulation | watch |
| [Human-as-Humanoid](https://zgc-embodyai.github.io/Human-as-Humanoid) | 2026-06 | arXiv | Synchronized ego-exo human videos converted through motion recovery, IK, and FK-aware supervision into 60-DoF humanoid action chunks | Zero-shot humanoid VLA learning from human demonstrations | watch |
| [UMI-Bench 1.0](https://arxiv.org/abs/2606.10382) | 2026-06 | arXiv | Local-first real-robot evaluation protocol dedicated to UMI-style data collection, deployment, reset, logging, and task-factor analysis | Reproducible UMI policy benchmarking | watch |
| [VISTA UMI](https://arxiv.org/abs/2606.04708) | 2026-06 | arXiv | Adapts UMI data for VLA training with wrist-fisheye UMI-VQA visual grounding plus physics-based trajectory validation | Making UMI demonstrations usable for VLAs | watch |
| [Pose6DAug](https://arxiv.org/abs/2606.20118) | 2026-06 | arXiv | Physically plausible multi-view object-swapping augmentation for VLA policies under occlusion and egocentric viewpoints | Robust manipulation data augmentation | watch |
| [MotionWAM](https://arxiv.org/abs/2606.09215) | 2026-06 | arXiv | Real-time humanoid world-action model driven by a single egocentric camera and unified whole-body motion tokens | Humanoid loco-manipulation from egocentric perception | watch |
| [OSCAR](https://arxiv.org/abs/2606.04463) | 2026-06 | arXiv | Omni-embodiment action-conditioned video world model trained across robotics and egocentric human datasets | Robot policy evaluation and cross-embodiment world modeling | watch |
| [Mem-World](https://arxiv.org/abs/2606.18960) | 2026-06 | arXiv | Wrist-view-centered 4D surfel memory for persistent action-conditioned manipulation world models | Long-horizon wrist-camera world modeling | watch |
| [PAIWorld](https://arxiv.org/abs/2606.18375) | 2026-06 | arXiv | 3D-consistent multi-view world foundation model across egocentric, eye-to-hand, and wrist cameras | Multi-view robotic world modeling | watch |
| [Qwen-RobotManip](https://arxiv.org/abs/2606.17846) | 2026-06 | arXiv | 38,100-hour manipulation pretraining corpus converting egocentric hand demonstrations into robot trajectories across 15 platforms | Scalable VLA robot manipulation | watch |
| [CAIP](https://arxiv.org/abs/2606.17256) | 2026-06 | arXiv | Contrastive action-image pretraining from 32,041 hours of egocentric human video with 3D hand-keypoint action proxies | Action-centric visual pretraining for robotics | watch |
| [YUBI](https://arxiv.org/abs/2606.10244) | 2026-06 | arXiv | Reported 8,434 hours, 1.20M episodes, and 119 tasks from a finger-aligned UMI-style bimanual interface | Scalable bimanual dexterous data collection | watch |
| [HumanoidUMI](https://arxiv.org/abs/2606.27239) | 2026-06 | arXiv | VR-assisted egocentric data capture for whole-body humanoid imitation, using sparse human keypoint trajectories and wrist-view observations for whole-body transfer | Whole-body humanoid transfer from egocentric human demonstrations | watch |
| [BRIDGE State-Gated Experts](https://arxiv.org/abs/2606.26603) | 2026-06 | arXiv | Combines handheld UMI demonstrations with targeted teleoperated segments through state-gated diffusion-policy experts | Contact-rich manipulation with mixed UMI and teleoperation supervision | watch |
| [ForceBand](https://forceband-emg.github.io) | 2026-06 | arXiv | 10-hour wrist-worn sEMG, egocentric video, IMU, and fingertip-force dataset for force-enriched demonstrations | Force-aware human demonstration collection for robot policies | watch |
| [EgoEngine](https://egoengine.github.io/) | 2026-06 | arXiv | Converts egocentric human manipulation videos into high-fidelity robot observation videos and executable robot action trajectories | Human-to-robot data generation, dexterous imitation | watch |
| [EgoAERO](https://arxiv.org/abs/2606.08057) | 2026-06 | arXiv | Asset-free conversion from a single egocentric RGB-D demonstration; introduces EgoDex-R in the paper | Single-demo dexterous robot learning | watch |
| [1M-HUGs / HUG](https://grasping.io/) | 2026-06 | arXiv | 1M frames / 27.8 hours of smart-glasses human grasps over 6,707 object instances, plus HUG-Bench with 90 unseen objects | Human grasp modeling and zero-shot robot grasping | open |
| [HALOMI](https://arxiv.org/abs/2606.18772) | 2026-06 | arXiv | Extends UMI-style demonstration collection with egocentric head/wrist observations and head-hand trajectories for humanoid loco-manipulation | Active-perception humanoid manipulation from human demonstrations | watch |
| [HumanoidArena](https://arxiv.org/abs/2606.17833) | 2026-06 | arXiv | Benchmark for egocentric hierarchical whole-body learning across seven leg-critical humanoid-object and humanoid-scene interaction tasks | Whole-body humanoid control from egocentric perception | watch |
| [EgoStation Smartphone Raw Catalog v1](https://huggingface.co/datasets/zeno-labs/egostation-smartphone-raw-v1) | 2026-05 | Hugging Face | Metadata for 5,962 episodes and 693.9 hours, with 16 downloadable full-quality samples carrying raw video, hand keypoints, and depth overlays | Inspecting smartphone capture quality and requesting selected raw demonstrations | partial |
| [Luel IMU Egocentric 250k](https://huggingface.co/datasets/Luel-ai/imu-egocentric-250k) | 2026-05 | Hugging Face | Commercial card reports 250K hours and 46,340 tasks; the gated Hub repository exposes a verifiable 10-clip, 30-minute sample plus the full task taxonomy | Reviewing industrial RGB-IMU capture scope and requesting the proprietary full corpus | request |
| [RoboX-EgoTask](https://huggingface.co/datasets/RoboXTechnologies/RoboX-EgoTask) | 2026-05 | Hugging Face | 17 public RGB-D task clips from five recordings with hand signals, camera pose, IMU, action segments, trajectories, segmentation, and quality reports | Open multimodal hand-object interaction and human-demonstration prototyping | open |
| [SenseXperience UMI Human Demonstrations](https://huggingface.co/datasets/ICRA-WBCD/IO-AI-SenseXperience-UMI) | 2026-05 | Hugging Face | 1,051 public human-demonstration episodes, 220,963 frames, and 92 tasks with six synchronized head, wrist, depth, and gripper views plus 7-DoF poses in LeRobot format | Open human-to-robot transfer, bimanual manipulation, and VLA data | open |
| [tau0-WM](https://arxiv.org/abs/2606.01027) | 2026-05 | arXiv | Unified video-action world model trained on real-robot teleoperation, UMI-style interaction, egocentric human video, and rollout/failure data | Future-aware robot action generation and action-conditioned simulation | watch |
| [SABER](https://arxiv.org/abs/2605.09613) | 2026-05 | arXiv | 100+ hours of in-store retail egocentric action capture paired with 360-degree exocentric context | Retail-domain VLA adaptation and manipulation data scaling | watch |
| [Qwen-VLA](https://arxiv.org/abs/2605.30280) | 2026-05 | arXiv | Unified VLA model trained over manipulation trajectories, egocentric human demonstrations, simulation, navigation, and trajectory supervision | General embodied action modeling | watch |
| [Gaze2Act](https://arxiv.org/abs/2605.30282) | 2026-05 | arXiv | Maps first-person gaze into robot perspective for target specification in interactive manipulation | Gaze-conditioned VLA manipulation | watch |
| [DeMiAn](https://arxiv.org/abs/2605.17077) | 2026-05 | arXiv | Dense multi-aspect language re-annotation over 1M robot clips and 50K EgoVerse human-egocentric videos | Language-dense robot policy learning | watch |
| [Grid Egocentric Residential Samples](https://huggingface.co/datasets/aryan56789gridai/grid-egocentric-residential-samples) | 2026-05 | Hugging Face | Public first-person residential manipulation preview pack with full-length 1080p household task videos | Household VLA capture-quality review and imitation-learning samples | open |
| [Industrial Workplace Egocentric FHD Samples](https://huggingface.co/datasets/TrainThemAI/Industrial-Workplace-Egocentric-FHD-Samples) | 2026-05 | Hugging Face | 21 rights-cleared head-mounted workplace clips with paired JSON metadata across factory, construction, services, retail, transit, and food-service domains | Industrial and vocational human-demonstration samples for VLA/WAM prototyping | open |
| [UniDataPro Egocentric Video Preview](https://huggingface.co/datasets/UniDataPro/egocentric-video) | 2026-05 | Hugging Face | Commercial preview card for a claimed 4,050-hour VR-headset/Zed first-person robotics collection; public files expose a sample video, CSV, and tracking text | Market scan and modality reference for large-scale egocentric robotics capture | partial |
| [Lo6yu Egocentric RGB-D + EMG/IMU](https://huggingface.co/datasets/Lo6yu/egocentric_dataset) | 2026-05 | Hugging Face | Eight household activity packages with RGB-D, wrist EMG/IMU, hand keypoints, masks, contact, per-finger force, semantic segments, and training-ready exports | Multimodal contact-aware imitation learning and sensor-fusion robot data | open |
| [Egocentric Adjust Bottle LeRobot](https://huggingface.co/datasets/suz22/egocentric_adjust_bottle) | 2026-05 | Hugging Face | LeRobot-format human_mano adjust-bottle task with 500 episodes and 36,897 frames across image, text, timeseries, and video exports | Loader tests and small-scale LeRobot manipulation policy prototyping | open |
| [Low-Resolution Active Perception BC](https://arxiv.org/abs/2605.14106) | 2026-05 | arXiv | Behavior cloning with low-resolution wrist-mounted egocentric RGB for object finding and grasp triggering | Active perception under low compute | watch |
| [EgoSPT / SPOT](https://arxiv.org/abs/2605.20085) | 2026-05 | arXiv | Egocentric spatially prompted manipulation trajectories with first-frame object/target grounding and 3D end-effector motion | Spatially grounded manipulation trajectory prediction | watch |
| [Mobile UMI](https://arxiv.org/abs/2605.20894) | 2026-05 | arXiv | Robot-free mobile manipulation capture with chest-centric context, wrist-centric interaction, decoupled kinematics, and latency-aware diffusion policy | Mobile manipulation from portable demonstrations | watch |
| [BifrostUMI](https://arxiv.org/abs/2605.03452) | 2026-05 | arXiv | VR-device robot-free demonstrations with sparse keypoint trajectories, wrist-mounted visual data, and humanoid retargeting | Humanoid whole-body demonstration collection | watch |
| [UniT](https://arxiv.org/abs/2604.19734) | 2026-04 | arXiv | Unified latent action tokenizer anchoring massive egocentric human data to humanoid policy learning and world modeling | Human-to-humanoid transfer | watch |
| [XRZero-G0](https://github.com/X-Square-Robot/XRZero-G0) | 2026-04 | arXiv | VR and dual-gripper robot-free collection system with a reported 2,000-hour dataset and data-mixing study | Scalable robot-free manipulation data collection | watch |
| [EgoVerse](https://egoverse.ai/) | 2026-04 | arXiv | 1,362 hours, 80K episodes, 1,965 tasks, 240 scenes, 2,087 demonstrators | Human demonstration scaling for robot learning and VLA | watch |
| [RoboX Egocentric Collection](https://huggingface.co/datasets/RoboXTechnologies/RoboX-Egocentric-Collection-v0.2) | 2026-04 | Hugging Face | 7,342 auto-gated first-person clips across grasping, daily activities, scene capture, and navigation with metadata, hand keypoints, object tracks, action segments, IMU, and camera pose | Multi-task robotics imitation data and navigation/manipulation benchmark prototyping | request |
| [EgoLive](https://arxiv.org/abs/2604.23570) | 2026-04 | arXiv | Large-scale real-world task-oriented egocentric routines for robot manipulation | Home service, retail, and real-world work-task manipulation | watch |
| [GazeVLA](https://gazevla.github.io/) | 2026-04 | arXiv | VLA policy that pretrains on large-scale egocentric human data to capture gaze, intention, and action before robot fine-tuning | Gaze- and intent-aware robot manipulation | watch |
| [WARPED](https://arxiv.org/abs/2604.10809) | 2026-04 | arXiv | Wrist-aligned rendering converts monocular egocentric human demonstrations into robot policy observations with 3D Gaussian Splatting | Cross-embodiment imitation from human ego video | watch |
| [UMI-3D](https://umi-3d.github.io/) | 2026-04 | arXiv | Extends UMI with wrist LiDAR, LiDAR-centric SLAM, synchronized sensing, and spatiotemporal calibration for robust 3D demonstration capture | 3D-aware UMI data collection | watch |
| [OmniUMI](https://arxiv.org/abs/2604.10647) | 2026-04 | arXiv | Adds RGB, depth, trajectory, tactile sensing, grasp force, and external wrench signals to a human-aligned UMI-style handheld system | Contact-rich multimodal UMI demonstrations | watch |
| [HRDexDB](https://arxiv.org/abs/2604.14944) | 2026-04 | arXiv | 1.4K human/robot grasping trials, tactile, multiview video, egocentric video streams | Cross-domain dexterous grasp learning | watch |
| [UMI-Underwater](https://umi-under-water.github.io/) | 2026-03 | arXiv | Transfers on-land handheld human demonstrations to underwater grasping through depth-based affordance representations | Underwater manipulation without underwater teleoperation | watch |
| [HoMMI](https://hommi-robot.github.io/) | 2026-03 | arXiv | Whole-body mobile manipulation interface augmenting UMI with egocentric sensing, relaxed head actions, and cross-embodiment hand-eye policy design | Robot-free mobile manipulation demonstrations | watch |
| [Ego-Exo Manufacturing](https://huggingface.co/datasets/third-origin/ego-exo-manufacturing) | 2026-03 | Hugging Face | Gated 275-hour shoe-manufacturing corpus with 40 synchronized ego-exo groups, 80 standalone ego sessions, 137 exo sessions, and procedure/proficiency/mistake annotations | Industrial ego-exo learning, skill assessment, and procedural modeling | request |
| [OBayData Egocentric Dexterous Manipulation Demo](https://huggingface.co/datasets/obaydata/egocentric-dexterous-manipulation-demo) | 2026-03 | Hugging Face | 500 train and 100 test first-person manipulation sessions with head and wrist videos, 3D hand joints, camera extrinsics, action labels, and language annotations | Quality review and prototyping for dexterous human-demonstration pipelines | open |
| [HuMI](https://humanoid-manipulation-interface.github.io/) | 2026-02 | arXiv | Portable robot-free whole-body demonstration interface with open code and seven Hugging Face datasets across kneeling, squatting, tossing, walking, and bimanual tasks | Humanoid whole-body manipulation from portable human demonstrations | open |
| [Seesaw Video Subset](https://huggingface.co/datasets/SeesawVideos/Video_Subset) | 2026-02 | Hugging Face | 763 gated LeRobot-compatible indoor episodes, including 243 ego and 520 exo views, 321,178 frames, four room settings, and smartphone AR camera pose | Ego-exo robot learning, action recognition, and data-conversion experiments | request |
| [CoMe-VLA](https://arxiv.org/abs/2602.04600) | 2026-02 | arXiv | Cognitive and memory-aware VLA that learns active-perception strategies from large-scale egocentric human data | Non-Markovian active perception and manipulation | watch |
| [EgoAVFlow](https://arxiv.org/abs/2602.22461) | 2026-02 | arXiv | Learns manipulation and active camera control from egocentric human videos through shared 3D flow | Active-vision robot policy transfer | watch |
| [EgoHumanoid](https://arxiv.org/abs/2602.10106) | 2026-02 | arXiv | Co-trains humanoid VLA policies from robot-free egocentric human demonstrations and limited robot data | Humanoid loco-manipulation | watch |
| [EgoActor](https://arxiv.org/abs/2602.04515) | 2026-02 | arXiv | Grounds high-level instructions into spatially aware egocentric humanoid actions | Humanoid task planning and action grounding | watch |
| [AoE: Always-on Egocentric](https://arxiv.org/abs/2602.23893) | 2026-02 | arXiv | Always-on egocentric human-video collection pipeline and corpus for embodied AI | Scaling human-video data for robot learning | watch |
| [EgoScale](https://arxiv.org/abs/2602.16710) | 2026-02 | arXiv | 20,854 hours of action-labeled egocentric human video with a human-to-robot two-stage transfer recipe and a log-linear data-scaling law | Scaling dexterous manipulation from human video | watch |
| [Manus Egocentric Sample](https://huggingface.co/datasets/OpenGraphLabs-Research/manus-egocentric-sample) | 2026-01 | Hugging Face | 20 LeRobot-format manipulation episodes with egocentric RGB/depth, Manus glove tracking, IMU, timeseries, and videos | Open egocentric manipulation loader tests and sensor-fusion policy prototyping | open |
| [HoyerChou Egocentric Bimanual Videos](https://huggingface.co/datasets/HoyerChou/EgocentricVideos) | 2026-01 | Hugging Face | YOTO/YOTO++ release with three raw first-person bimanual videos, 119 segmented demos, 241 rows, MANO hand meshes, 3D keypoints, calibration, and camera/world transforms | One-shot bimanual manipulation and hand-trajectory extraction from human video | open |
| [Egocentric Specialties](https://huggingface.co/datasets/ThisIsPepper/EgocentricSpecialties) | 2026-01 | Hugging Face | Auto-gated skilled-jewelry manufacturing data card describing synchronized egocentric and top-down footage for diamond setting and specialty workflows | Skilled industrial manipulation and fine-grained manufacturing demonstrations | request |
| [X-Humanoid](https://arxiv.org/abs/2512.04537) | 2025-12 | arXiv | Applies human-to-humanoid video translation to 60 hours of Ego-Exo4D, releasing 3.6M robotized humanoid frames | Humanoid video/world-model data generation | watch |
| [DreamTacVLA](https://arxiv.org/abs/2512.23864) | 2025-12 | arXiv | Grounds VLA policies in contact physics using tactile images, wrist-camera local vision, and third-person macro vision | Contact-rich VLA manipulation | watch |
| [TacThru-UMI](https://arxiv.org/abs/2512.09851) | 2025-12 | arXiv | See-through-skin tactile-visual sensing plus a UMI-style imitation-learning framework for multimodal manipulation | Tactile-visual robot policy learning | watch |
| [Mitty](https://arxiv.org/abs/2512.17253) | 2025-12 | arXiv | Diffusion Transformer for translating human demonstrations into robot-execution videos, with synthesis from large egocentric datasets | Human-to-robot video generation | watch |
| [World in Your Hands / WiYH](https://arxiv.org/abs/2512.24310) | 2025-12 | arXiv | Reported 1,000-hour egocentric manipulation ecosystem with multiview/depth/hand/wrist signals | VLA, manipulation representation learning | watch |
| [Hoi!](https://arxiv.org/abs/2512.04884) | 2025-12 | arXiv | 3,048 force-grounded articulated-manipulation sequences over human, wrist-camera, UMI, and Hoi-gripper embodiments | Force/contact-grounded cross-view manipulation | watch |
| [METIS](https://arxiv.org/abs/2511.17366) | 2025-11 | arXiv | Dexterous VLA pretrained on multi-source egocentric human and robotic data unified under one action space | Dexterous egocentric-VLA pretraining | watch |
| [Let Me Show You / RfV](https://arxiv.org/abs/2511.05199) | 2025-11 | IROS 2025 | Retrieves egocentric human demonstrations and extracts affordance masks plus hand trajectories for robot policies | Retrieval-from-video robot manipulation | watch |
| [UMIGen](https://arxiv.org/abs/2511.09302) | 2025-11 | arXiv | Cloud-UMI plus visibility-aware egocentric point-cloud generation | 3D-aware cross-embodiment imitation learning | watch |
| [EgoMI](https://arxiv.org/abs/2511.00153) | 2025-10 | arXiv | Egocentric manipulation interface capturing synchronized end-effector and active-head trajectories for semi-humanoid robots | Active-vision whole-body manipulation | watch |
| [FARM](https://tactile-farm.github.io) | 2025-10 | arXiv | Force-aware diffusion policy using a modified tactile UMI gripper and matched actuated deployment gripper | Force- and tactile-grounded manipulation | watch |
| [UMI-on-Air](https://arxiv.org/abs/2510.02614) | 2025-10 | arXiv | Embodiment-aware diffusion policy adapting handheld UMI demonstrations to aerial manipulators | UMI transfer under embodiment dynamics | watch |
| [EmbodiSwap](https://arxiv.org/abs/2510.03706) | 2025-10 | arXiv | Photorealistic robot overlays over ego-centric human video for synthetic robot imitation datasets | Zero-shot human-video-to-robot imitation | watch |
| [Hand-VLA Pretraining](https://arxiv.org/abs/2510.21571) | 2025-10 | arXiv | Converts unscripted real-life egocentric hand videos into 1M VLA-aligned action-language episodes | Robot pretraining from human activity video | watch |
| [ActiveUMI](https://arxiv.org/abs/2510.01607) | 2025-10 | arXiv | Portable VR/UMI-style system capturing active egocentric perception for bimanual manipulation | Active-perception UMI-style data collection | watch |
| [FastUMI-100K](https://huggingface.co/datasets/IPEC-COMMUNITY/FastUMI_100k_lerobot) | 2025-10 | arXiv | 100K+ UMI-style trajectories across 54 tasks in LeRobot format with multi-view wrist fisheye images | Large-scale UMI-style robot policy training | open |
| [BiNoMaP](https://hnuzhy.github.io/projects/BiNoMaP/) | 2025-09 | arXiv | Category-level bimanual non-prehensile primitives learned from egocentric hand trajectories and transferred across two dual-arm platforms | Contact-rich bimanual skill extraction from human demonstrations | watch |
| [EMMA](https://arxiv.org/abs/2509.04443) | 2025-09 | IEEE RA-L 2025 | Co-trains egocentric human mobile-manipulation data with static robot data to avoid mobile teleoperation bottlenecks | Scalable mobile manipulation from human data | watch |
| [OpenEgo](https://arxiv.org/abs/2509.05513) | 2025-09 | arXiv | 1,107 hours unified across six public egocentric datasets, 290 manipulation tasks, hand-pose/action primitives | Standardized dexterous manipulation pretraining and evaluation | watch |
| [MV-UMI](https://arxiv.org/abs/2509.18757) | 2025-09 | arXiv | Multi-view UMI adds third-person context to wrist egocentric observations | Broad-scene context for cross-embodiment manipulation | watch |
| [InterVLA](https://arxiv.org/abs/2508.04681) | 2025-08 | ICCV 2025 | 11.4 hours and 1.2M frames with 2 ego and 5 exo views, human/object motions, and verbal commands | Vision-language-action and motion estimation | watch |
| [EgoVLA](https://rchalyang.github.io/EgoVLA/) | 2025-07 | arXiv | VLA training from egocentric human videos plus robot fine-tuning | Human-video-to-robot policy transfer | watch |
| [EgoDex](https://arxiv.org/abs/2505.11709) | 2025-05 | ICLR 2026 | 829 hours of Apple Vision Pro egocentric video, 194 tabletop tasks, 3D hand/finger tracking | Dexterous manipulation, imitation learning, human-to-robot hand trajectory prediction | watch |
| [TASTE-Rob](https://arxiv.org/abs/2503.11423) | 2025-03 | CVPR 2025 | 100,856 task-oriented ego-centric hand-object interaction videos aligned with language instructions | Video generation and robot imitation learning | watch |
| [Kaiwu](https://arxiv.org/abs/2503.05231) | 2025-03 | IEEE RA-L 2025 | 11,664 synchronized assembling-action instances with first-person video, gaze, hand motion, pressure, audio, EMG, mocap, and robot streams | Multimodal robot learning and HRI | watch |
| [Humanoid-VLA](https://arxiv.org/abs/2502.14795) | 2025-02 | arXiv | Combines language, egocentric scene perception, and motion control for universal humanoid control | Humanoid VLA and motion generation | watch |
| [Immersive Social Interaction with VR and LLM-Assisted Humanoids](https://arxiv.org/abs/2607.07430) | 2024-11 | Humanoids 2024 Workshop | Apple Vision Pro teleoperation with egocentric robot feedback, voice locomotion, wrist/finger retargeting, and RGB, gaze, command, joint-state, and hand-motion recording on a Unitree H1 | Immersive humanoid teleoperation and multimodal demonstration capture | watch |
| [EgoMimic](https://arxiv.org/abs/2410.24221) | 2024-10 | ICRA 2025 | Project Aria human videos with 3D hand tracking and robot co-training for unified imitation learning | Human-to-robot imitation from ego video | watch |
| [FastUMI](https://arxiv.org/abs/2409.19499) | 2024-09 | arXiv | UMI redesign reporting 10K+ real-world trajectories across 22 everyday tasks | Faster hardware-independent UMI-style collection | watch |
| [R+X](https://arxiv.org/abs/2407.12957) | 2024-07 | ICRA 2025 | Retrieves executable skills from long unlabeled first-person videos of everyday human tasks | Robot imitation from everyday first-person video | watch |
| [UMI on Legs](https://umi-on-legs.github.io) | 2024-07 | arXiv | Combines handheld UMI task demonstrations with simulation-trained whole-body controllers for quadruped manipulation | Mobile cross-embodiment UMI transfer | watch |
| [EgoPAT3D / EgoPAT3Dv2](https://arxiv.org/abs/2403.05046) | 2024-03 | ICRA 2024 | 1M+ RGB-D/IMU frames in the original task; v2 expands egocentric 3D action-target prediction | Human-robot interaction, 3D target anticipation, manipulation safety | open |
| [Universal Manipulation Interface / UMI](https://umi-gripper.github.io/) | 2024-02 | RSS 2024 | Hand-held GoPro gripper interface and policy stack for in-the-wild robot teaching | Wrist-view robot-free demonstrations and cross-embodiment policy transfer | open |
| [Dobb-E / HoNY](https://dobb-e.com/) | 2023-11 | arXiv | 13 hours, 22 homes, 5,620 trajectories, and 1.5M RGB-D frames collected with an iPhone-based Stick | Household robot learning from handheld in-home demonstrations | partial |
| [VIP](https://arxiv.org/abs/2210.00030) | 2022-09 | ICLR 2023 | Value-Implicit Pre-Training learns visual rewards and representations from Ego4D human video for few-shot robot manipulation | Ego4D-to-robot reward learning | open |
| [R3M](https://arxiv.org/abs/2203.12601) | 2022-03 | CoRL 2022 | Robot-manipulation visual representation pretrained on Ego4D with time-contrastive and video-language objectives | Ego4D-to-robot representation learning | open |

### Hands, Objects, Pose, Tracking, and 3D

Fine-grained hand, object, contact, and 3D-pose datasets, including emerging event-camera and EMG sensing.

### Daily Life, Memory, Assistance, and QA

Long-horizon daily-life capture and the question-answering benchmarks that probe memory, reasoning, and assistance.

### Action, Procedure, Lifelogging, and Classic FPV

Procedural and activity datasets, from modern industrial assembly to the classic first-person-vision benchmarks that started the field.

### Project Aria, AR/VR, and 3D Scene Resources

AR-glasses and headset data with gaze, SLAM, and digital-twin annotations for scene-level perception.

## Benchmarks and Derived Annotations

Evaluation suites and label sets built on top of the raw datasets above.

## Models, Tools, and Baselines

Open models, baselines, and loaders you can build on directly.

### Egocentric Foundation Models and Assistants

### Video-Language and Long-Video Models

### Action, Tracking, Pose, and HOI Baselines

### Practical Tooling

## Adjacent and Related Resources

These resources sit outside the atlas's human-worn egocentric core, but provide useful comparisons and transfer paths: robot-mounted or simulated first-person views, general human video, multi-view robotics, autonomous driving, and long-video reasoning. They are tagged `scope: adjacent` in [`data/resources.yml`](data/resources.yml) so public counts and filters keep the distinction explicit.

| Resource | Released | Venue | Scale / signal | Why it is adjacent (not egocentric) | Status |
| :--- | :---: | :---: | :--- | :--- | :---: |
| [AlloEgo-VLM](https://github.com/CKL9001/AlloEgo-VLM) | 2026-08 | arXiv | Public AlloEgo-View triplets and SFT code disambiguate allocentric versus egocentric spatial references, including an Isaac Sim object-search test | General images and simulation rather than wearable capture; repository-wide license is absent | partial |
| [WorldRover-10M](https://arxiv.org/abs/2608.15659) | 2026-08 | arXiv | Synthetic minute-scale exploration from first-, third-, and 360-degree views with RGB, depth, poses, actions, flow, and long-range tracks | Artist-built synthetic worlds; code and data links are not yet live | watch |
| [EATR-Stereo](https://arxiv.org/abs/2608.17453) | 2026-08 | arXiv | Embodiment-conditioned routing of paired head-stereo evidence reports 60% full-task and 100% grasp success on a 33-DoF humanoid | Robot head cameras rather than human wearable video | watch |
| [Hydra-0](https://nvidia-isaac.github.io/video_to_data/hydra-0/) | 2026-08 | arXiv | Action-flow world model trained after filtering 2,202 hours of multi-embodiment video, supporting replay evaluation and human-demo transfer | General robot/human video; code and models are still coming soon | watch |
| [DECOWAM](https://arxiv.org/abs/2608.20114) | 2026-08 | arXiv | Separates camera ego-motion, base, and arm action in a whole-body world-action model and introduces the ARMDOG real-robot dataset | Moving robot-camera perception; artifacts are not yet released | watch |
| [RoboEdit](https://arxiv.org/abs/2608.18948) | 2026-08 | arXiv | Turns human manipulation video into aligned robot video and 3D states; reports RoboEdit-14M with 174K pairs across seven embodiments | Human video need not be wearable first-person; data and code remain unreleased | watch |
| [HiPHI](https://noitom-robotics.github.io/hiphi/) | 2026-08 | Hugging Face | Gated 617.5-hour optical-motion corpus with 200.1M frames, 132 performers, and 245.7 hours of hand-object interaction | Full-body and HOI mocap without a central first-person camera | request |
| [NVIDIA Video-to-Data Robot Dexterity Dataset](https://huggingface.co/datasets/nvidia/video-to-data-robot-dexterity-task-library-and-dataset) | 2026-08 | Hugging Face | 1,215 TACO-derived motions retargeted into 121,500 bimanual robot episodes totaling 924.8 GB, with CC-BY-4.0 data and public tools | Synthetic robot executions derived from existing human demonstrations | open |
| [RoboLab-EgoX](https://huggingface.co/datasets/Voxel51/RoboLab-EgoX) | 2026-08 | Hugging Face | 4,000 simulation episodes over 28 tasks and 99 scenes, including success/failure labels, wrist RGB-D, exterior RGB-D, telemetry, and actions | Robot wrist-view simulation benchmark rather than human wearable capture | open |
| [OccPlanner](https://arxiv.org/abs/2608.14160) | 2026-08 | arXiv | Occupancy-conditioned pixel-goal diffusion planner with L3ROcc supervision from monocular robot videos; reports 71.55% average success at 5-8 m versus 20.81% for NavDP | Robot-view navigation rather than human wearable capture | watch |
| [AdvDex](https://arxiv.org/abs/2608.14028) | 2026-08 | arXiv | Aligns human and robot dexterous demonstrations through an SE(3) wrist plus 15-joint hand action space, tactile/kinematic OmniShare data, and adversarial embodiment learning | OmniShare is not established as a wearable first-person corpus | watch |
| [ContactGuard](https://arxiv.org/abs/2608.13438) | 2026-08 | arXiv | Predicts action-conditioned visual latents and aborts likely manipulation failures before contact from wrist and multi-view robot cameras | Robot execution monitoring rather than human wearable capture | watch |
| [HumanoidVLN](https://humanoid-vln.github.io/) | 2026-08 | arXiv | Isaac Sim benchmark with four humanoids, 933 collision-aware routes, four instructions per route, and a 20-episode Unitree G1 sim-to-real pilot | Simulated humanoid first-person observations rather than human wearable capture | watch |
| [Proxemic Risk from Robot Ego Images](https://arxiv.org/abs/2608.12515) | 2026-08 | ECCV 2026 EMR Workshop | Four-level danger classification study of InternVL, Qwen-VL, and SmolVLM with prompting, QLoRA, and person-localization analysis | Robot viewpoint and safety labels rather than human wearable capture | watch |
| [Fast-Slow ReAct ObjectNav](https://arxiv.org/abs/2608.09816) | 2026-08 | arXiv | Bounded VLM deliberation over coordinate-anchored semantic memory and pose-tagged first-person keyframes; reports 68.75% HM3D and 47.29% MP3D success | Navigation-robot observations rather than human wearable capture | watch |
| [VideoNIG](https://arxiv.org/abs/2608.08596) | 2026-08 | arXiv | 60K simulated ego-view tours and 37K multimodal prompts for navigation-instruction generation, spatial diagnosis, and downstream execution | Simulator-generated agent views rather than human wearable capture | watch |
| [WNM-3D](https://arxiv.org/abs/2608.07267) | 2026-08 | arXiv | Conditions joint future-view and action generation on geometry tokens reconstructed from monocular navigation history | Navigation-agent camera history rather than human wearable capture | watch |
| [AtlasVLA](https://arxiv.org/abs/2608.06729) | 2026-08 | arXiv | Wrist-camera VLA with persistent voxel-hashed 4D world memory and task-progress-aware ego working memory | Robot wrist-camera manipulation rather than human wearable capture | watch |
| [CrossTracer](https://arxiv.org/abs/2608.06688) | 2026-08 | arXiv | Cross-embodiment image-plane trace prediction and residual adaptation, evaluated on NaviTrace and wheeled/legged robots | Robot-mounted egocentric sensing rather than human wearable capture | watch |
| [Omega-0](https://arxiv.org/abs/2608.06375) | 2026-08 | arXiv | Whole-body latent world-action model supporting robot ego RGB; reports Omega-HOME with 40+ hours of synchronized household humanoid observations and actions | Robot-mounted egocentric sensing rather than human wearable capture | watch |
| [LAWM-3D](https://arxiv.org/abs/2608.05706) | 2026-08 | arXiv | Learns 3D-aware latent actions from multiview human video for generalizable robot world models | Source human video is not confirmed as wearable first-person capture | watch |
| [HOST](https://host-site.host-robotics.workers.dev/) | 2026-07 | arXiv | One-video skill acquisition in 29 seconds on average with 62% reported success; training code and a 15.9 GB policy checkpoint are public | Human video is central but not required to be wearable first-person | partial |
| [Data Pyramid for Embodied Manipulation](https://jasper-aaa.github.io/embodied-data-pyramid/) | 2026-07 | arXiv | Open survey and living catalog organizing real-robot, UMI, human ego/ego-exo, simulation, and general video data into five levels | Broad embodied-manipulation taxonomy in which egocentric data is one major layer | open |
| [ContactFlow](https://arxiv.org/abs/2607.26579) | 2026-07 | arXiv | Embodiment-agnostic world-model conditioning through trajectories of 3D actor-object contact points, trained with human and robot interaction video | General human/robot video; wearable capture is not required | watch |
| [Robot-Factored World Models](https://bjkim95.github.io/rofacto/) | 2026-07 | arXiv | Factors actions into nominal robot trajectories and rendered geometry across DROID, egocentric RoboCasa-GR1, unseen robots, and retargeted human demos | Robot-camera world model rather than human wearable capture | watch |
| [AWM Navigation Pretraining Corpus](https://huggingface.co/datasets/Yangyihui/awm-nav-pretrain) | 2026-07 | Hugging Face | Public 198.1 GB corpus with 23,076 indoor/outdoor robot-view clips and paired Gemini-generated spatial/action labels; source-derived reuse terms are incomplete | Robot navigation observations rather than human wearable capture | partial |
| [Think at 5 Hz, Act at 20 Hz](https://arxiv.org/abs/2607.15621) | 2026-07 | arXiv | Asynchronous driving VLA with a frozen 7B reasoner and fast action expert, raising CARLA route completion from 37.0 to 94.0 at fresh 20 Hz control | Autonomous-driving camera input rather than human wearable capture | watch |
| [IMBench](https://imbench.org/) | 2026-07 | RSS 2026 SemRob Workshop | Open benchmark with 35 tasks, seven categories, and 14K trajectories integrating physical reasoning with executable manipulation | Robot-manipulation benchmark rather than human wearable capture | open |
| [AC-VLA](https://arxiv.org/abs/2607.15714) | 2026-07 | arXiv | Compositional VLA training with instruction decomposition, trajectory alignment, and wrist-view masking; reports about 28% absolute LIBERO-OOD improvement | Robot VLA method rather than a human first-person resource | watch |
| [SkillNav](https://arxiv.org/abs/2607.15758) | 2026-07 | arXiv | Training-free object-goal navigation skills written into a curiosity map, reaching 43.2 SPL and 75.9% success on HM3D v0.2 | Robot navigation observations rather than human wearable capture | watch |
| [Orbis 2](https://lmb-freiburg.github.io/orbis2.github.io/) | 2026-07 | arXiv | Hierarchical driving world model that separates long-horizon scene structure from detailed generation; code and checkpoints remain coming soon | Autonomous-driving world model rather than human wearable capture | watch |
| [Data and Learning Where it Matters](https://arxiv.org/abs/2607.15982) | 2026-07 | arXiv | Targeted collection and offline RL for contact-critical segments averages 96% success across four real tasks from 2-2.5 hours of data | Robot contact-rich manipulation rather than human wearable capture | watch |
| [MotionForesight](https://motionforesight.github.io/) | 2026-07 | arXiv | Forecasts future object-centered 3D scene flow by repurposing frozen video and tracking priors trained on 40K monocular human-object videos | General monocular interaction video without a wearable-capture requirement | watch |
| [VTM-Nav](https://arxiv.org/abs/2607.14514) | 2026-07 | arXiv | Training-free object-goal navigation with persistent visual-topological room/object memory, coarse-to-fine retrieval, and a conservative execution guard | Robot navigation memory rather than human wearable capture | watch |
| [SafeRelBench](https://arxiv.org/abs/2607.14543) | 2026-07 | arXiv | 507 household-agent evaluations testing support, containment, and proximity constraints before risk-prone actions | Robot-agent spatial-safety evaluation rather than human wearable capture | watch |
| [SoftNav](https://arxiv.org/abs/2607.14586) | 2026-07 | IROS 2026 | Continuous 3D soft tokens ground a frozen VLM in detected objects and frontiers using about 1,200 samples and 17M trainable parameters | Robot object-goal navigation rather than human wearable capture | watch |
| [Representation-Aligned Tactile Grounding](https://arxiv.org/abs/2607.14609) | 2026-07 | arXiv | Latent tactile predictor aligns intermediate VLA features with future contact consequences without reconstructing noisy raw tactile signals | Robot tactile-VLA learning rather than human wearable capture | watch |
| [Action QFormer](https://arxiv.org/abs/2607.14635) | 2026-07 | arXiv | Instruction-conditioned query interface reorganizes inherited multimodal features before action generation, with large reported sim-to-real gains | Robot navigation VLA architecture rather than human wearable capture | watch |
| [Reflex](https://arxiv.org/abs/2607.14695) | 2026-07 | ICML 2026 | Streaming flow-matching VLA runtime reports 2.58x speedup, stable 50 Hz control, and up to 54% lower reaction latency | Robot VLA inference framework rather than human wearable capture | watch |
| [FoMoVLA](https://liauto-research.github.io/FoMoVLA/) | 2026-07 | arXiv | Future-feature foresight and sparse 2D trajectories jointly supervise target states and motion paths for continuous policies | Robot VLA foresight method rather than human wearable capture | watch |
| [LifelongVLA](https://arxiv.org/abs/2607.14852) | 2026-07 | arXiv | Continual VLA learning separates reusable shared knowledge from task-specific experts while controlling forgetting | Robot continual-learning method rather than human wearable capture | watch |
| [Steering Robustness into World Action Models](https://arxiv.org/abs/2607.14943) | 2026-07 | arXiv | Robustness-oriented WAM training uses uncertainty-aware perturbations and corrective trajectories to improve recovery under distribution shift | Robot world-action-model training rather than human wearable capture | watch |
| [AeroAct](https://arxiv.org/abs/2607.14997) | 2026-07 | arXiv | Vision-language-action policy couples aerial perception, language-conditioned planning, and low-level flight control | Aerial-robot VLA rather than human wearable capture | watch |
| [DriftWorld](https://driftworld.github.io/) | 2026-07 | arXiv | Driving world model explicitly represents scene dynamics and controllable ego motion for long-horizon generation | Autonomous-driving world modeling rather than human wearable capture | watch |
| [BadWAM](https://arxiv.org/abs/2607.15207) | 2026-07 | arXiv | Benchmark and attack framework studies backdoors in world-action models across visual triggers and action outcomes | Robot WAM security rather than human wearable capture | watch |
| [RoboTTT](https://research.nvidia.com/labs/gear/robott/) | 2026-07 | arXiv | One-shot human-video imitation combines online test-time training with world-model supervision, reporting five-minute learning of a 10-stage assembly task | Human-video robot adaptation without a released wearable-video corpus | watch |
| [PhysClaw-0](https://open-gigaai.github.io/PhysClaw/) | 2026-07 | arXiv | Open collect-verify-reset system that stores and reuses language corrections, reducing human working time to 16% and raising single-attempt success from 12.5% to 47.5% | Robot autonomous-data-collection toolkit rather than human wearable capture | open |
| [Industrial Dexterity Benchmark](https://arxiv.org/abs/2607.14021) | 2026-07 | arXiv | Industrial hardware boards plus DAG-ROS and AG-iDP3 for cable, harness, and gearbox tasks; best cable setup reaches 78% versus 36% for single-camera RGB | Industrial robot benchmark rather than human wearable capture | watch |
| [GigaWorld-Policy-0.5](https://open-gigaai.github.io/giga-world-policy/) | 2026-07 | arXiv | Open action-centered WAM with mixed world/action training, MoT experts, AutoResearch, and 85 ms RTX 4090 inference without future-video decoding | Robot world-action model rather than human wearable video | open |
| [VSI-Super-Wild](https://vsi-super-wild.github.io/) | 2026-07 | ECCV 2026 | 6,980 human-verified spatial QA pairs over 442 continuous panoramic videos totaling 284.52 hours; public 2.67 TB data lacks explicit reuse terms | Panoramic YouTube-derived video rather than wearer capture | partial |
| [REAL / REAL-Bench](https://github.com/InternRobotics/REAL) | 2026-07 | ECCV 2026 | Open mobile-manipulation framework and 241-task benchmark reporting 56.9% interactive and 78.3% physical-robot success, with code and 3.92 GB of data | Robot-only open-world manipulation rather than human wearable capture | open |
| [UESF-Bench / SeekFollow-VLA](https://arxiv.org/abs/2607.13621) | 2026-07 | arXiv | Unified language-guided human seeking and following with semantic exploration, delayed identity grounding, phase switching, and recovery | Robot-agent human search/following rather than human wearable video | watch |
| [JOP-VLN](https://qingrongh.github.io/JOP-VLN) | 2026-07 | IROS 2026 | Joint imitation, DAgger, and reinforcement learning reaches 69.9% success on R2R and 68.0% on RxR | Navigation-agent observations rather than human wearable capture | watch |
| [Reverse to Advance](https://arxiv.org/abs/2607.13455) | 2026-07 | arXiv | Trains difficult manipulation from easier reversed rollouts using automated collection, temporal inversion, kinematic filtering, critic filtering, and iterative refinement | Robot manipulation data collection rather than human first-person data | watch |
| [Anchor-Align](https://anchoralignvla.github.io/) | 2026-07 | arXiv | Representation anchoring plus language-action alignment improves physical xArm success from 28% to 54% and 37% to 60%; linked code currently returns 404 | Robot-camera VLA finetuning rather than human wearable capture | watch |
| [JITOMA / JITOMA-Bench](https://arxiv.org/abs/2607.13245) | 2026-07 | arXiv | Just-in-time scene-graph growth activates task-relevant memory anchors to curb long-horizon perceptual saturation, graph size, and caption latency | Robot scene-graph memory rather than human wearable video | watch |
| [WANDA / Worlds in One Demo](https://wanda.lecar-lab.org/) | 2026-07 | arXiv | Public 16,922-episode, roughly 248-hour synthetic mobile-manipulation set expands one real demonstration per task across generated 3D worlds | Synthetic robot trajectories rather than human wearable capture | open |
| [GPUSimBench](https://arxiv.org/abs/2607.13059) | 2026-07 | IROS 2026 | Tests Isaac Lab and Genesis for real-to-sim physical consistency, parallel throughput and memory, plus run-to-run and inter-environment nondeterminism | Robot-simulator infrastructure benchmark without human capture | watch |
| [HRIBench (Interaction-Centric VLA)](https://arxiv.org/abs/2607.13056) | 2026-07 | arXiv | 13 role-conditioned tasks and 650+ episodes assess intent, synchronization, responsiveness, protocol, and safety across instructor, collaborator, and intruder roles | Robot-view collaboration benchmark rather than human wearable capture | watch |
| [DenseReward](https://arxiv.org/abs/2607.13033) | 2026-07 | arXiv | Dense frame-level vision-language rewards trained with automatically synthesized physical failures for manipulation, MPC, and RL; the claimed release page currently returns 404 | Robot-camera reward learning rather than human wearable capture | watch |
| [FlowWAM](https://flow-wam.github.io/) | 2026-07 | arXiv | Open optical-flow world-action model reporting 92.94%/92.14% RoboTwin Clean/Random success and a 63.71 WorldArena EWMScore, with code, checkpoints, and flow data | Robot-only manipulation and world modeling rather than human wearable capture | open |
| [ChunkFlow](https://arxiv.org/abs/2607.12992) | 2026-07 | arXiv | Continuity-consistent chunked-policy learning with editable overlap zones, seam and derivative losses, robust history training, and AWAC adaptation | Robot-policy stability method without a human first-person resource | watch |
| [ExToken](https://arxiv.org/abs/2607.12931) | 2026-07 | arXiv | Reinforcement fine-tuning that conditions VLA exploration on demonstration-derived behavioral tokens and learns a deployment-time token selector | Robot VLA exploration method rather than egocentric human data | watch |
| [Hy-Embodied-VLM-1.0](https://github.com/Tencent-Hunyuan/HY-Embodied) | 2026-07 | arXiv | Open approximately 30B MoE embodied VLM activating about 3B parameters per token; evaluated on 38 benchmarks with first-place results on 19 among compared similarly sized models | General physical-world agent model rather than a human wearable resource | open |
| [Jetson-PI](https://github.com/PKU-SEC-Lab/Jetson-PI) | 2026-07 | arXiv | Open foresight-aligned asynchronous VLA stack reporting 8.66x higher control frequency than naive PyTorch, 5.41x over vla.cpp, and +14.8% average success over VLASH | Onboard robot-control runtime rather than human first-person capture | open |
| [TrustVLA](https://arxiv.org/abs/2607.12571) | 2026-07 | arXiv | Retraining-free VLA backdoor defense using evidence-evolution monitoring, counterfactual trigger localization, and localized inpainting | Robot VLA security method rather than an egocentric data resource | watch |
| [Self-in-Space / SIS-Bench](https://choucisan.github.io/publications/self-in-space/) | 2026-07 | ACM MM 2026 | Open 4,856-question benchmark over 1,646 UAV videos and 13 tasks, plus 54,298 training examples, code, and a motion-aware adapter | Agent-centered UAV onboard video rather than human wearable capture | open |
| [VistaVLA](https://arxiv.org/abs/2607.12356) | 2026-07 | arXiv | 3D-Gaussian-grounded VLA reporting 99% context-token reduction, +22.8% success across seven real tasks, and +30.0% OOD over VLA-Adapter | Multi-view robot manipulation rather than human wearable video | watch |
| [Efficient VLA Inference via Temporal Redundancy Reduction](https://arxiv.org/abs/2607.12287) | 2026-07 | arXiv | Dynamic visual-token updates plus two-step diffusion action sampling report more than 2x speedup and up to 98% success across LIBERO, RoboTwin, and real robots | Robot VLA inference optimization rather than a first-person resource | watch |
| [REGRIND](https://yunhaifeng.com/REGRIND) | 2026-07 | arXiv | Single-demonstration retargeting plus residual RL for scissors and screwdriver use on LEAP and WUJI hands; linked code repository is currently unavailable | Human hand-object references and robot control, but no wearable-video release | watch |
| [E-VQA / ST-Evidence](https://github.com/SalesforceAIResearch/EVQA) | 2026-07 | ECCV 2026 | Open benchmark for answers with temporal segments and tracked object masklets, plus 160K instruction examples and released 3B/7B models | General video QA and evidence grounding rather than egocentric video | open |
| [Xiaomi-Robotics-U0](https://arxiv.org/abs/2607.11643) | 2026-07 | arXiv | 38B embodied-synthesis model for multi-view scenes, embodiment transfer, editing, and video; the claimed release URL currently returns 404 | Robot and embodied-scene generation rather than human wearable capture | watch |
| [World Action Models to Embodied Brains Roadmap](https://arxiv.org/abs/2607.11689) | 2026-07 | arXiv | Roadmap spanning WAM representations, standardization, physical harnesses, shared contracts, and closed-loop post-training | Broad physical-intelligence position paper rather than an egocentric resource | watch |
| [DA-Nav / ReDA](https://arxiv.org/abs/2607.11638) | 2026-07 | arXiv | Direction-aware city-scale VLN and recovery data, reporting 56.16% unseen-city success and kilometer-scale robot trials | Robot-agent egocentric navigation, not human wearable capture | watch |
| [Robot-Centric Pointmaps](https://davian-robotics.github.io/pointmap/) | 2026-07 | arXiv | Robot-frame per-pixel XYZ improves pi0.5 by 7.6 points and SmolVLA by 4.2 on RoboCasa, with larger gains under unseen real-camera viewpoints | Robot-camera VLA representation rather than human first-person data | watch |
| [TeleDexter](https://bigai-dex.github.io/blog/teledexter/) | 2026-07 | arXiv | Hand-object co-tracking controller reporting 75.2% average success across seven tasks and two dexterous hands, with demonstrations used for behavior cloning | Robot teleoperation from human references, not a released wearable-video resource | watch |
| [WALA](https://liujiahao2077.github.io/WALA.github.io) | 2026-07 | arXiv | Executable latent actions from labeled demonstrations and action-free videos using DINOv3 feature deltas, depth, and latent world modeling | General robot-policy learning; code is still marked coming soon and no egocentric corpus is released | watch |
| [SLVMBench](https://arxiv.org/abs/2607.11312) | 2026-07 | arXiv | 2,261 questions over 1,220 videos test learning a tutorial embedded in a 2-3-hour stream and applying it to an ongoing task | General long-video skill memory rather than first-person-centered data | watch |
| [Lumo-2](https://arxiv.org/abs/2607.11270) | 2026-07 | arXiv | Latent WAM with staged action-dynamics-vision-language alignment for scalable, OOD, long-horizon, and dexterous robot learning | Robot-centered WAM/VLA method without a wearable-video release | watch |
| [AdvNav](https://arxiv.org/abs/2607.11063) | 2026-07 | ACM MM 2026 | Black-box adversarial attack on first-person VLN observations with 49.70%-87.30% reported attack success against HAMT and MapGPT | Navigation-agent views rather than human wearable video | watch |
| [yyyyywv/egocentric](https://huggingface.co/datasets/yyyyywv/egocentric) | 2026-07 | Hugging Face | 32.8 GB, 1,019 robot episodes, and 994,459 frames across seven task groups from eye/head and wrist cameras; downloadable but the card omits provenance and capture documentation | Robot-mounted cameras and action/state streams rather than human wearable capture | partial |
| [TactiDex](https://tactidex.github.io/) | 2026-07 | arXiv | Human-demonstration benchmark aligning whole-hand tactile pressure, hand kinematics, object 6D states, language, and task phases, plus tactile-guided single/bimanual robot transfer | Tactile-glove and motion-capture demonstrations, not human wearable video | watch |
| [DemoBridge](https://gitlab.kuleuven.be/u0123974/demo-bridge/) | 2026-07 | RSS 2026 RoboData Workshop | Apache-2.0 toolkit that converts single-view stereo hand demonstrations into collision-aware, physics-validated robot trajectories | Human-hand retargeting without explicitly egocentric capture | open |
| [PhysV2A](https://arxiv.org/abs/2607.09365) | 2026-07 | arXiv | Converts video-derived 6D object motion into feasible robot trajectories using grasp-conditioned reachability checks and semantic-mask-constrained refinement | Generic video-to-robot transfer rather than human wearable capture | watch |
| [ACE-Brain-0.5](https://huggingface.co/ACE-Brain/ACE-Brain-0.5-8B) | 2026-07 | arXiv | Public Qwen3-VL-derived 8B-class checkpoint unifying spatial perception, planning, robot action generation, progress monitoring, and self-improvement across 15 benchmarks; no model license is declared | Robot-centric embodied foundation model with egocentric spatial reasoning, not human wearable capture | partial |
| [DexVerse](https://ycyao216.github.io/DexVerse.site/) | 2026-07 | arXiv | 100-task dexterous-manipulation benchmark across three arms and six hands, with VR teleoperation and 3,180 reported demonstrations; core BSD-3-Clause suite is live while full data/tooling remain pending | Robot simulation and teleoperation benchmark, not a human wearable-video corpus | partial |
| [FabriVLA](https://arxiv.org/abs/2607.08575) | 2026-07 | arXiv | Compact 1B VLA with InternVL3.5 and a flow-matching action head, reporting 90.0% tier-average success on Meta-World MT50 | Simulation-focused robot VLA, not human egocentric capture | watch |
| [Harness VLA](https://arxiv.org/abs/2607.08448) | 2026-07 | arXiv | Memory-guided agentic harness composing a frozen VLA with retryable contact primitives and fixed analytic primitives across LIBERO-Pro, RoboCasa365, and RoboTwin C2R | Robot VLA planning harness, not a first-person resource | watch |
| [AnyDexRT](https://chenxi-wang.github.io/projects/anydexrt/) | 2026-07 | arXiv | Calibration-free hand retargeting with self-supervised fingertip correspondence, few-shot human guidance, and contact-aware pinch refinement | Human-hand teleoperation method, not egocentric video data | watch |
| [TFP](https://arxiv.org/abs/2607.08283) | 2026-07 | RSS 2026 SemRob Workshop | Event-sensitive memory-action fusion for a 3.3B VLA, evaluated on LIBERO, LIBERO-plus, and MIKASA ShellGameTouch | Robot VLA memory method without wearable capture | watch |
| [LingBot-VA 2.0](https://arxiv.org/abs/2607.08639) | 2026-07 | arXiv | Native video-action model with a semantic tokenizer, causal pretraining, sparse MoE backbone, and asynchronous closed-loop inference | Robot video-action foundation model, not human first-person data | watch |
| [LEEVLA](https://arxiv.org/abs/2607.08182) | 2026-07 | arXiv | Task-aware VLA using drift-guided visual prioritization and structured latent feature-flow generation; the linked repository does not yet match the paper | Robot VLA architecture without human wearable capture | watch |
| [Video-Action Generalization Gap / Temporal Ratio](https://umishra.me/temporal-ratio/) | 2026-07 | arXiv | Diagnostic Temporal Ratio and adaptive inference guidance for compositional generalization in video-action models on LIBERO and real robot tasks | Robot video-action analysis, not a human first-person resource | watch |
| [GIRAF](https://arxiv.org/abs/2607.07880) | 2026-07 | CVPR 2026 HuMoGen Workshop | Text-conditioned full-body interaction generation with articulated objects, object-centric contact representations, and contact augmentation | Generated third-person human-object motion, not wearable capture | watch |
| [LingBot-World 2.0](https://arxiv.org/abs/2607.07534) | 2026-07 | arXiv | Interactive world simulator with unbounded interaction horizon, 720p/60 FPS real-time distillation, richer action/event controls, and 14B/1.3B models | General interactive world simulator, not human first-person capture | watch |
| [RoboDojo](https://arxiv.org/abs/2607.04434) | 2026-07 | arXiv | Unified sim-and-real robot manipulation benchmark with 42 simulation tasks, 18 real tasks, remote evaluation, 30 integrated policies, and a leaderboard | Robot-policy benchmark, not human wearable capture | watch |
| [LaMem-VLA](https://arxiv.org/abs/2607.07608) | 2026-07 | arXiv | Latent-memory-native VLA framework that reconstructs short- and long-term history into latent memory tokens for long-horizon manipulation | Robot VLA memory architecture, not a human first-person resource | watch |
| [TouchWorld](https://arxiv.org/abs/2607.07287) | 2026-07 | arXiv | Predictive-and-reactive tactile foundation model separating vision-language planning, tactile world prediction, action generation, and residual refinement | Robot tactile manipulation method, not wearable capture | watch |
| [WAM-TTT](https://arxiv.org/abs/2607.06988) | 2026-07 | arXiv | Test-time training framework steering frozen WAMs from unlabeled human videos through adaptive memory and paired human-robot meta-training | Human-video-to-robot WAM method, not a human wearable dataset | watch |
| [EAGOR](https://arxiv.org/abs/2607.06165) | 2026-07 | arXiv | Training-free spherical-belief framework for 360-degree directional reasoning on HOS, OSR-Bench, and legged-robot navigation | Embodied 360-camera robot/agent reasoning, not human wearable capture | watch |
| [VLA Models Review](https://arxiv.org/abs/2607.06706) | 2026-07 | arXiv | Survey of 183 VLA contributions from 2017-2026 across architectures, training recipes, bimanual coordination, UAVs, memory, and world models | Robot/UAV VLA survey, not an egocentric resource | watch |
| [NativeMEM](https://arxiv.org/abs/2607.06678) | 2026-07 | arXiv | VLA memory compression that turns each historical camera frame into a single memory token for low-latency long-horizon manipulation | Robot VLA memory method, not human wearable capture | watch |
| [Lift3D-VLA](https://arxiv.org/abs/2607.06564) | 2026-07 | arXiv | 3D geometry- and dynamics-aware VLA with point-cloud reasoning, geometry-centric masked autoencoding, and 22 sim / 8 real tasks | Robot-only 3D VLA method, not human first-person data | watch |
| [Pelican-VLA 0.5](https://arxiv.org/abs/2607.06655) | 2026-07 | arXiv | Unified VLA report combining vision-language understanding, future-frame generation, action prediction, and reasoning-slot attention | Robot VLA architecture, not human wearable capture | watch |
| [ActionCache](https://arxiv.org/abs/2607.06370) | 2026-07 | arXiv | Training-free cache/refinement method accelerating flow-matching VLA action heads by reusing intermediate actions with compact multimodal keys | Robot VLA inference method, not a first-person resource | watch |
| [Smooth Operator / SBR](https://arxiv.org/abs/2607.07491) | 2026-07 | arXiv | Sampling-based hand retargeter for low-jitter real-time teleoperation, evaluated in simulation and an 18-participant manipulation study | Teleoperation/retargeting method, not a first-person dataset | watch |
| [LongVQUBench](https://arxiv.org/abs/2607.01086) | 2026-07 | ECCV 2026 | 1,200+ long videos and 1,500 questions for video-quality reasoning, spanning movies, documentaries, surveillance, egocentric recordings, and animation | Broad long-video quality benchmark with egocentric recordings as one slice, not centered on wearable capture | watch |
| [TAP VLA](https://arxiv.org/abs/2607.02466) | 2026-07 | ICML 2026 | Task-agnostic VLA pretraining learns motor priors from unlabeled robot interaction data before language grounding | Robot/VLA pretraining, not human wearable capture | watch |
| [The Moving Eye](https://arxiv.org/abs/2607.02322) | 2026-07 | IROS 2026 | Hybrid dynamic data collection with one robot arm acting as a moving environmental camera while the other manipulates | Robot camera/viewpoint generalization, not human first-person capture | watch |
| [Bridge-WA](https://hcplab-sysu.github.io/BRIDGE-WA) | 2026-07 | arXiv | World-action framework distilling future-change tokens, change maps, and motion-flow maps into a VLA action transformer | Robot world-action model, not wearable capture | watch |
| [VLA-Corrector](https://arxiv.org/abs/2607.01804) | 2026-07 | arXiv | Lightweight detect-and-correct inference for action-chunked VLA policies using visual feature evolution | Robot VLA correction method, not human wearable capture | watch |
| [Embodied.cpp](https://arxiv.org/abs/2607.02501) | 2026-07 | arXiv | Portable C++ runtime for VLA/WAM inference on heterogeneous robots with adapters, head plugins, and robot deployment layers | Embodied robot runtime, not a human first-person resource | watch |
| [EVA-Client](https://arxiv.org/abs/2607.02646) | 2026-07 | arXiv | Unified real-robot client for policy deployment, data collection, evaluation, rollout logging, and comparison views | Robot policy client and evaluator, not wearable capture | watch |
| [InternVLA-A1.5](https://arxiv.org/abs/2607.04988) | 2026-07 | arXiv | Robot VLA with native VLM backbone, continuous action expert, subtask/VQA supervision, and latent foresight tokens | Robot-only VLA architecture, not human first-person data | watch |
| [SIEVE](https://arxiv.org/abs/2607.06442) | 2026-07 | arXiv | Structure-aware selection of VLA demonstrations by reusable primitives, composition patterns, and medoid trajectories | Robot demonstration-selection method, not egocentric capture | watch |
| [SVA / Look Before You Leap](https://arxiv.org/abs/2607.03751) | 2026-07 | arXiv | Distills tree-search rollouts into an action evaluator for frozen VLA policies at test time | Robot VLA action-evaluation method, not wearable capture | watch |
| [WorldOdysseyBench](https://arxiv.org/abs/2606.31672) | 2026-06 | arXiv | 600+ first/third-person interactive world-model test cases with 10-60s WASD interaction and action, vision, physics, and memory metrics | General interactive world-model benchmark, not human wearable capture | watch |
| [Qwen-RobotNav](https://arxiv.org/abs/2606.18112) | 2026-06 | arXiv | Qwen-branded navigation model trained on 15.6M samples with task modes and controllable visual-history observation parameters | Robot/agent navigation model, not human wearable capture | watch |
| [Qwen-RobotWorld](https://arxiv.org/abs/2606.17030) | 2026-06 | arXiv | Language-conditioned embodied video world model with an 8.6M video-text EWK corpus over 20+ embodiments and 500+ action categories | Robot, navigation, driving, and human-to-robot world modeling rather than human first-person capture | watch |
| [FutureNav](https://arxiv.org/abs/2606.30367) | 2026-06 | arXiv | Unified world-action modeling for vision-and-language navigation with action, dynamics, and future spatial-state objectives | Navigation-agent WAM, not human wearable capture | watch |
| [ZR-0](https://github.com/RUCKBReasoning/ZR-0) | 2026-06 | arXiv | 2.6B VLA trained with dense embodied chain-of-thought supervision over ProcCorpus-60M for cross-embodiment manipulation | Robot VLA pretraining, not a human first-person dataset | watch |
| [VLK](https://vision-language-kinematics.github.io/) | 2026-06 | arXiv | 48K synthetic humanoid loco-manipulation trajectories with rendered egocentric observations, instructions, and whole-body kinematics | Synthetic humanoid supervision, not real human wearable capture | watch |
| [HUMEMBR](https://arxiv.org/abs/2606.30404) | 2026-06 | IROS 2026 | Human-centered long-term memory for embodied robot QA and routine-conditioned navigation | Robot memory/navigation system, not an egocentric capture resource | watch |
| [RhinoVLA](https://arxiv.org/abs/2606.07383) | 2026-06 | arXiv | Edge-deployable VLA using a token-efficient Qwen3-VL backbone, Action Expert, View Registry, and hardware-aware execution | Robot-only VLA deployment system, not wearable capture | watch |
| [OpenEAI-Platform](https://arxiv.org/abs/2606.03392) | 2026-06 | arXiv | OpenEAI-Arm plus OpenEAI-VLA built on Qwen3-VL-4B and a Diffusion Transformer action head | Robot-arm hardware/software platform, not human first-person data | watch |
| [SAGE-Nav](https://arxiv.org/abs/2606.25497) | 2026-06 | arXiv | LLM-planned object-goal navigation with dynamic scene graphs and egocentric robot observations | Robot object-navigation method, not human wearable capture | watch |
| [USS](https://arescheah.github.io/uss-project-page/) | 2026-06 | arXiv | Unified spatial-semantic prompting for embodied visual tracking with text, point, box, and mask prompts | Robot/agent tracking method, not human first-person capture | watch |
| [RoboAtlas](https://arxiv.org/abs/2606.26046) | 2026-06 | arXiv | Contextual Active SLAM with OpenRoboVox semantic mapping and egocentric VLM reasoning | Robot active-SLAM/navigation system, not human wearable capture | watch |
| [iCrowdNav](https://arxiv.org/abs/2606.26047) | 2026-06 | arXiv | Intention-aware crowd navigation from egocentric robot observations and human-pose interaction encoding | Robot crowd-navigation method, not human wearable capture | watch |
| [Cloak](https://arxiv.org/abs/2606.22836) | 2026-06 | arXiv | Masks the robot end-effector from wrist-camera VLA observations for zero-shot cross-embodiment transfer | Wrist-camera robot VLA method, not human first-person capture | watch |
| [OpenHLM](https://openhlm-project.github.io/) | 2026-06 | arXiv | Whole-body humanoid loco-manipulation recipe with teleoperation, VLA design, and HuMI co-training | Robot-centered humanoid VLA recipe, not human wearable capture | watch |
| [UniviewVLA](https://arxiv.org/abs/2606.21501) | 2026-06 | arXiv | Multiview VLA with world modeling for occlusion-aware robot manipulation from agent and wrist cameras | Robot VLA/world-modeling method, not human first-person capture | watch |
| [ContactWorld](https://arxiv.org/abs/2606.13877) | 2026-06 | arXiv | Vision-tactile world-model benchmark over 12 contact-rich manipulation tasks with wrist/front views, point clouds, and tactile force fields | Robot contact-rich manipulation benchmark, not human wearable capture | watch |
| [3DG-VLN / UAV-VLN-FOV](https://github.com/xuefanfu/3DG-VLN) | 2026-06 | arXiv | 2,717 target-visible UAV trajectories with high-resolution egocentric observations and 3D waypoint labels | Aerial-agent navigation benchmark, not human wearable capture | watch |
| [NavWAM](https://arxiv.org/abs/2606.13494) | 2026-06 | arXiv | Navigation world-action model predicting future egocentric visual observations and action sequences for goal-conditioned visual navigation | Robot visual-navigation world model, not human wearable capture | watch |
| [IntentNav](https://arxiv.org/abs/2606.08029) | 2026-06 | arXiv | Spatial-visual object-navigation policy learned from human demonstrations with frontier probing and spatial memory | Object-navigation work from human demonstrations, not a human wearable resource | watch |
| [AlloSpatial](https://arxiv.org/abs/2606.08952) | 2026-06 | arXiv | Agentic spatial-reasoning harness that converts local egocentric observations into allocentric spatial representations | General spatial-reasoning framework, not a wearable capture dataset | watch |
| [World-Language-Action Model](https://arxiv.org/abs/2606.05979) | 2026-06 | arXiv | Unified model predicting subtasks, subgoal images, and robot actions while learning from egocentric videos and robot trajectories | Broader embodied model using egocentric video among robot data, not centered on human wearable capture | watch |
| [ManiSplat](https://arxiv.org/abs/2606.10645) | 2026-06 | arXiv | Manipulation trajectory synthesis from monocular video using decoupled 3D Gaussian Splatting | Monocular manipulation synthesis; first-person human capture not confirmed | watch |
| [DexFuture](https://arxiv.org/abs/2606.05699) | 2026-06 | arXiv | Hierarchical future-state visuomotor targeting for bimanual dexterous tool use | Robot dexterity method, not human first-person data | watch |
| [GRAIL](https://arxiv.org/abs/2606.05160) | 2026-06 | arXiv | Digital-twin pipeline generating humanoid loco-manipulation demonstrations from 3D assets and video priors | Humanoid demonstration generation, not a human wearable dataset | watch |
| [DataLadder](https://arxiv.org/abs/2606.16776) | 2026-06 | arXiv | Simulation-enabled interconversion toolchain for robot trials, demonstrations, synthetic data, and evaluation assets | Embodied-data tooling rather than egocentric capture | watch |
| [EgoInfinity](https://huggingface.co/datasets/Rice-RobotPI-Lab/egoinfinity) | 2026-06 | arXiv | Public 46 GB preview with 104 Action100M clips and retargeted trajectories for four robot embodiments; the full web-scale engine is not released | Human-video-to-robot data engine, not centered on wearable first-person capture | partial |
| [LUCID](https://lucid-robot.github.io/) | 2026-06 | arXiv | Embodiment-agnostic intent model learned from unstructured human videos, paired with simulation-trained dexterous robot control | Unstructured human-video-to-robot transfer; source videos are not confirmed as egocentric wearable capture | watch |
| [Dexterous Point Policy](https://arxiv.org/abs/2606.10614) | 2026-06 | arXiv | Transfers raw human videos into dexterous hand policies through task-relevant 3D keypoints for hands and objects | Human demonstration video transfer, not confirmed first-person capture | watch |
| [TopoRetarget](https://arxiv.org/abs/2606.16272) | 2026-06 | arXiv | Interaction-preserving retargeting from human hand-object demonstrations to dexterous robot-hand references | Human hand-object retargeting method, not confirmed wearable first-person capture | watch |
| [WARP Retarget](https://warp-retarget.github.io/) | 2026-06 | arXiv | Whole-body-aware retargeting from offline human demonstrations into mobile-manipulator trajectories | Offline human-demonstration retargeting, not confirmed wearable first-person capture | watch |
| [V2P-Manip](https://arxiv.org/abs/2606.16436) | 2026-06 | arXiv | Learns dexterous manipulation from monocular human videos through 3D assets, trajectory estimation, physical refinement, and policy learning | Monocular human-video-to-robot work, not confirmed wearable capture | watch |
| [SeeTraceAct](https://arxiv.org/abs/2606.02745) | 2026-06 | arXiv | Demo-conditioned VLA with visibility-aware future end-effector traces and RoboCasa-DC cross-embodiment demonstration episodes | Robot VLA conditioned on demonstrations, not a human wearable dataset | watch |
| [What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos?](https://arxiv.org/abs/2606.06627) | 2026-06 | arXiv | 532 everyday human videos with 28 hours of high-quality hand labels for studying robot-policy cotraining under limited robot data | Everyday human-video transfer study, not scoped as egocentric capture | watch |
| [LARA](https://arxiv.org/abs/2606.07100) | 2026-06 | arXiv | Latent Action Representation Alignment for jointly optimizing latent action models and VLAs from unlabeled human videos | Human-video-to-VLA method, not itself an egocentric dataset | watch |
| [Video2Sim2Real](https://arxiv.org/abs/2606.08828) | 2026-06 | arXiv | Reconstructs a simulator-ready digital twin and robot/object motion priors from a single human manipulation video | Single-human-video robot skill acquisition, not confirmed first-person wearable data | watch |
| [HOWTransfer](https://arxiv.org/abs/2606.10743) | 2026-06 | arXiv | Localizes hand-object contact in human video demonstrations and retargets wrist trajectories into robot-executable motions | Hand-centric human-video-to-robot transfer, not confirmed egocentric wearable capture | watch |
| [Retrieve, Don't Retrain](https://arxiv.org/abs/2606.15631) | 2026-06 | arXiv | Retrieval-augmented VLA adaptation using pool-side demonstrations such as human-hand videos instead of per-task retraining | Robot VLA adaptation with human-hand video pools, not centered on first-person data | watch |
| [SERF](https://arxiv.org/abs/2606.12956) | 2026-06 | arXiv | Spatiotemporal environment and robot feature map updated from egocentric observations and proprioception for long-horizon mobile manipulation | Robot-centric policy work using egocentric observations, not human wearable capture | watch |
| [ImageWAM](https://arxiv.org/abs/2606.19531) | 2026-06 | arXiv | World-action model that uses image editing instead of dense video generation for robot action prediction | Robot WAM comparison point, not a human first-person resource | watch |
| [CHORD](https://nvidia-isaac.github.io/video_to_data/chord/) | 2026-06 | arXiv | NVIDIA contact-wrench-guided dexterous RL with 4,739 simulation tasks, 82.12% success over 1,831 evaluated tasks, whole-body transfer, and real-hardware results | Motion-capture and third-person human demonstrations rather than a wearable-video release; code remains forthcoming | watch |
| [JAXenstein](https://arxiv.org/abs/2605.19926) | 2026-05 | arXiv | JAX Wolfenstein 3D renderer for fast reinforcement-learning experiments in visual first-person tasks | Virtual first-person RL environment, not real egocentric capture | watch |
| [ACWM-Phys](https://arxiv.org/abs/2605.08567) | 2026-05 | arXiv | Controllable action-conditioned world-model benchmark for rigid, kinematic, deformable, and particle dynamics | Simulator benchmark for physical interaction, not human first-person capture | watch |
| [SpatialBench](https://ropedia.github.io/SpatialBench/) | 2026-05 | arXiv | Broad spatial foundation-model benchmark across viewpoints, scene domains, input densities, and hardware constraints | Useful context for egocentric spatial reasoning, but not centered on human wearable capture | watch |
| [LEXI-SG](https://ori-drs.github.io/lexisg-web/) | 2026-05 | arXiv | Monocular RGB open-vocabulary 3D scene-graph mapping validated on self-collected egocentric office sequences | Robot/agent mapping work, not centered on human wearable capture | watch |
| [EgoTraj-Bench](https://zoeyliu1999.github.io/EgoTrajBench/) | 2026-04 | ICRA 2026 | 36,947 public sequences from 17 robot FPV sessions for converting noisy first-person trajectories into clean bird's-eye trajectories | Robot-mounted first-person views rather than human wearable capture | open |
| [EgoDyn-Bench](https://huggingface.co/datasets/TUM-AVS/EgoDyn-Bench-ECCV2026) | 2026-04 | ECCV 2026 | Public 1,000-clip physics-grounded driving benchmark with about 14K oracle-labelled QA pairs, dynamics arrays, and a 49-model reference leaderboard | Autonomous-driving ego-motion video rather than human or animal wearable capture | open |
| [DenseStep2M](https://huggingface.co/datasets/mingjige/DenseStep2M) | 2026-04 | arXiv | About 100K instructional videos and 2M dense procedural steps from automated long-video annotation | General instructional-video corpus with egocentric transfer tests, not first-person centered | watch |
| [StarVLA](https://arxiv.org/abs/2604.05014) | 2026-04 | arXiv | Modular VLA codebase supporting swappable VLM backbones such as Qwen-VL and world-model backbones such as Cosmos | VLA tooling for comparisons, not human first-person capture | watch |
| [Point of View Robot Sociability Study](https://arxiv.org/abs/2603.28272) | 2026-03 | arXiv | Immersive-VR study comparing allocentric, egocentric-proximal, and egocentric-distal ratings of identical robot trajectories and a head-nod signal | Simulated pedestrian viewpoints rather than recorded wearable video | watch |
| [RoboMirror](https://arxiv.org/abs/2512.23649) | 2025-12 | arXiv | Retargeting-free humanoid locomotion from raw egocentric or third-person video using VLM-derived motion intent and a diffusion policy | Supports egocentric input but targets humanoid locomotion | watch |
| [VLA-Arena](https://arxiv.org/abs/2512.22539) | 2025-12 | ICML 2026 | Open-source VLA benchmark with 170 structured tasks and decoupled task, language, and visual perturbation axes | Robot VLA benchmark, not a human wearable/egocentric dataset | watch |
| [VLSA / AEGIS](https://arxiv.org/abs/2512.11891) | 2025-12 | IROS 2026 | Plug-and-play safety constraint layer for VLA policies plus SafeLIBERO safety-critical manipulation benchmark | Robot VLA safety architecture, not human wearable capture | watch |
| [UniWM](https://github.com/UWMILab/UniWM) | 2025-10 | ECCV 2026 | Memory-augmented world model for navigation foresight and planning, with public code and a Hugging Face dataset | Robot navigation cameras rather than human wearable capture | partial |
| [MM-Nav](https://pku-epic.github.io/MM-Nav-Web/) | 2025-10 | arXiv | Multi-view VLA navigation model with 360-degree observations and synthetic expert data for reaching, squeezing, and avoiding | Robot/synthetic visual navigation rather than human wearable capture | watch |
| [Seeing Across Views / MV-RoboBench](https://github.com/microsoft/MV-RoboBench) | 2025-10 | ICLR 2026 | 1.7K curated QA items over eight subtasks for multi-view spatial reasoning of VLMs in robotic manipulation (ICLR 2026) | Multi-camera robot scenes, not wearable capture | open |
| [HUI360](https://hucebot.github.io/hui360/) | 2025-09 | FG 2026 | 99 mobile-robot 360-degree HRI recordings, a 1M-annotation HUI360 open set, 6M SSUP-HRI annotations, and public anticipation baselines | Captured from a mobile robot, not a human camera wearer | partial |
| [HomeSafeBench](https://github.com/BITHLP/HomeSafeBench) | 2025-09 | arXiv | Open VirtualHome benchmark with 1,000 test tasks, 3,400 training tasks, five hazard classes, and 3,158 CueBack trajectories | Rendered embodied-agent observations rather than wearable capture | open |
| [RoboPearls](https://arxiv.org/abs/2506.22756) | 2025-06 | arXiv | Editable 3D Gaussian video simulation for robot manipulation, evaluated on RLBench, COLOSSEUM, Ego4D, Open X-Embodiment, and real robots | Robot simulation/editing tool, not a human egocentric dataset | watch |
| [NORA](https://arxiv.org/abs/2504.19854) | 2025-04 | arXiv | 3B generalist VLA using Qwen2.5-VL-3B and 970K real-world robot demonstrations | Robot-only demonstrations and observations, not human wearable capture | watch |
| [SEED4D](https://seed4d.github.io/) | 2024-12 | WACV 2025 | Synthetic ego-exo dynamic 4D generator and autonomous-driving dataset (16.8M images, vehicle cameras, LiDAR; WACV 2025) | Vehicle-egocentric driving data, not human/wearable | open |
| [Open X-Embodiment / RT-X](https://robotics-transformer-x.github.io/) | 2023-10 | ICRA 2024 | 1M+ real robot trajectories, 22 robot embodiments, 60 pooled robot datasets, standardized RLDS; [unofficial Hugging Face mirror](https://huggingface.co/datasets/jxu124/OpenX-Embodiment) | Robot-mounted and wrist cameras, not human first-person; pairs with egocentric human-video VLA pretraining | open |

## Surveys, Papers, and Context

Foundational papers and recent surveys for background and citation.

| Resource | Why it matters |
| :--- | :--- |
| [Vision-Language Models for Egocentric Video](https://arxiv.org/abs/2608.18671) | Critical survey connecting hand-object interaction, temporal and graph reasoning, efficient long-video processing, wearable assistance, and embodied AI. |
| [Ego4D: Around the World in 3,000 Hours of Egocentric Video](https://arxiv.org/abs/2110.07058) | Defines the modern large-scale egocentric benchmark suite. |
| [Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives](https://arxiv.org/abs/2311.18259) | Defines the major synchronized ego-exo activity resource. |
| [Scaling Egocentric Vision: The EPIC-KITCHENS Dataset](https://arxiv.org/abs/1804.02748) and [EPIC-KITCHENS-100](https://arxiv.org/abs/2006.13256) | Core kitchen/action-recognition benchmark lineage. |
| [Egocentric Video-Language Pretraining](https://arxiv.org/abs/2206.01670) | Introduces EgoClip and EgoNCE, a central egocentric VLP recipe. |
| [EgoSchema](https://arxiv.org/abs/2308.09126) | A widely used diagnostic benchmark for very long-form video-language understanding. |
| [HoloAssist](https://arxiv.org/abs/2309.17024) | Important for human-AI assistance and instructor-performer interaction. |
| [Nymeria](https://arxiv.org/abs/2406.09905) | Major egocentric multimodal motion-language dataset from Project Aria. |
| [HOT3D](https://arxiv.org/abs/2406.09598) | Major 3D hand-object tracking dataset for AR/VR. |
| [OpenEgo](https://arxiv.org/abs/2509.05513), [EgoDex](https://arxiv.org/abs/2505.11709), [EgoVerse](https://arxiv.org/abs/2604.07607) | Recent large-scale egocentric manipulation data for robot learning. |
| [Challenges and Trends in Egocentric Vision: A Survey](https://arxiv.org/abs/2503.15275) | Recent comprehensive survey of egocentric tasks, datasets, and open problems. |
| [Bridging Perspectives: Cross-view Collaborative Intelligence with Egocentric-Exocentric Vision](https://arxiv.org/abs/2506.06253) | Survey of ego-exo collaboration and paired-capture research. |
| [World Action Models](https://arxiv.org/abs/2605.12090) | Systematizes WAMs across VLA, world models, portable human demonstrations, simulation, and internet-scale egocentric video. |

## Workshops and Challenges

Recurring venues and challenge hubs where new egocentric tasks and leaderboards appear.

| Event / hub | Focus |
| :--- | :--- |
| [SLT 2026 SmartGlasses Challenge](https://aslp-lab.github.io/SmartGlasses/) | Two-track evaluation of timestamped speaker-attributed ASR and spoken-language understanding over 106.98 hours and 714 four-channel smart-glasses audio sessions; data is distributed to registered participants. |
| [EgoLink 2026](https://ego-link.github.io/challenge2026/) | ACM MM 2026 grand challenge for egocentric social reasoning and interactive tool-using agents, with public labels, code, leaderboard, 4,030 MCQs, and 1,055 evaluation videos. |
| [Ego4D Challenges](https://ego4d-data.org/) | Episodic memory, forecasting, hand-object, social/audio, and other Ego4D tasks. |
| [Ego-Exo4D Challenges](https://ego-exo4d-data.org/) | Cross-view, correspondence, pose, skill, and proficiency tasks. |
| [EPIC-KITCHENS Challenges](https://epic-kitchens.github.io/) | Action recognition, action detection, anticipation, retrieval, and domain adaptation. |
| [HOT3D Challenge](https://facebookresearch.github.io/hot3d/) | 3D hand-object tracking and AR/VR pose estimation. |
| [Project Aria](https://www.projectaria.com/) | Aria datasets, tools, and AR sensing ecosystem. |
| [Joint Egocentric Vision (EgoVis) Workshop](https://egovis.github.io/cvpr26/) | Cross-dataset egocentric vision workshop (CVPR) spanning Ego4D, Ego-Exo4D, EPIC-KITCHENS, HoloAssist, and more. |
| [CVPR / ICCV / ECCV egocentric and embodied AI workshops](https://cvpr.thecvf.com/) | Search annually for "egocentric", "embodied", "ego-exo", "wearable", and "first-person". |

## Watchlist

These entries are promising but should be rechecked before treating them as stable public datasets in a paper or benchmark.

## Inclusion Rules

1. Prefer official project pages, GitHub repositories, Hugging Face datasets, arXiv pages, or university dataset pages.
2. Keep the public status conservative. If raw video is not clearly available, mark `partial` or `watch`.
3. Separate raw datasets from derived annotations and benchmarks.
4. Record modality, scale, task, license/access friction, and raw-video dependency when known.
5. Keep the main atlas strictly egocentric: human or animal first-person capture, or ego-exo resources where the ego view is central. List related but non-egocentric resources (robot-only datasets, multi-view robotic benchmarks, autonomous-driving data, general long-video reasoning) under [Adjacent and Related Resources](#adjacent-and-related-resources).
6. Keep pure dashcam/autonomous-driving and robot-embodiment data out of the main tables; surface it as context in Adjacent and Related Resources only when it is genuinely useful to egocentric work.
7. Keep figures text-light. Exact resource names and labels should remain in Markdown/SVG so readers can inspect them directly.

## Contributing

Contributions are welcome through pull requests and issues. Please include an official source, a concise description of what the resource contributes, and any known access or license notes.

See [`CONTRIBUTING.md`](CONTRIBUTING.md) for the inclusion policy and style. Automated checks run in CI to keep the README, catalog data, figures, and public exports consistent; contributors do not need to run local scripts before opening a PR.

## Cite This Atlas

If this atlas helps your research or project, a citation or a link back is appreciated. Machine-readable metadata lives in [`CITATION.cff`](CITATION.cff).

```bibtex
@misc{he_awesome_egocentric_atlas,
  author       = {He, Chaoyue},
  title        = {Awesome Egocentric Atlas: Egocentric AI Datasets, Benchmarks, Models, and Tools},
  year         = {2026},
  howpublished = {\url{https://github.com/ChaoYue0307/awesome-egocentric-atlas}}
}
```

awesome-egocentric-atlas

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

awesome-egocentric-atlas

오너의 다른 레포

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README