loopmaster

KO | EN

gitlite — search

#agent #java #linux #cpp #git #shell #ai-agents #typescript #reactjs #artificial-intelligence

(★ 9)

No description available.

파일 탐색기

ZIP 다운로드

- - plan.md
  - review.md
  - summary.md
  - trace.jsonl
- - plan.md
  - review.md
  - summary.md
  - trace.jsonl
- - plan.md
  - review.md
  - summary.md
  - trace.jsonl
- - plan.md
  - review.md
  - summary.md
  - trace.jsonl
- agentic-architecture.md
- - __init__.py
  - auditor.py
  - codex_subagent.py
  - handler.py
  - handler_chat.py
  - skill_updater.py
  - strategist.py
  - worker.py
  - workspace.py
- - arm_init_pose.json
  - head_camera_extrinsics.json
- - __init__.py
  - result.py
  - types.py
- - __init__.py
  - bridge.py
  - hei_rebot_lift_ik.py
  - mink_ik.py
  - solve_arm_ee_cli.py
- - __init__.py
  - base.py
  - dry_run.py
  - hei_rebot_lift.py
- - __init__.py
  - registry.py
- __init__.py
- __main__.py
- cli.py
- server_bridge.py
- setup_ik_conda_env.sh
- test_agentic_framework.py
- anygrasp_sdk
- Grounded-SAM-2
- - loopmaster-auto-demo.service
  - loopmaster-auto-demo.timer
- - app.py
  - README.md
- - auto_demo_order.py
  - backup_db.py
- - logo.png
  - loopmaster_logo.png
  - loopviz.css
  - style.css
- - about.html
  - admin_login.html
  - dashboard.html
  - index.html
  - loopviz.html
  - research.html
- - push_demo.py
  - test_demo_trajectory.py
  - test_pending_order_supersede.py
  - test_product_catalog.py
- app.py
- PROTOCOL.md
- README.md
- requirements.txt
.gitignore
.gitmodules
.python-version
capture_anygrasp_grasp_pose.py
hei-rebot-lift
keyboard_control.py
navigation_live_map.py
play_cache_traj_skill.py
pyproject.toml
README.md
requirements.txt
rerun_arm_camera_frames.py
run_chat_server.md
run_mujoco_ik.sh
run_telegrip.sh
uv.lock

# 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)

LoopMaster is a lightweight agentic robotics layer for the HEI ReBot
Lift real robot.

The handler-led subagent structure uses clear execution roles:

| LoopMaster role | Responsibility |
| --- | --- |
| Handler | Owns the run, workspace, robot connection, and role handoff. |
| Strategist | Selects registry-backed skills from the goal and writes `plan.md`. |
| Worker | Executes the plan, observes after control actions, and writes `summary.md`/`trace.jsonl`. |
| Auditor | Reviews trace evidence, detects missing task-level skills, and writes `review.md`. |

Motion runs must be closed-loop. A successful `action_sent` response only means
the command was accepted by the platform; agents must also check periodic or
post-action `observe` feedback and compare the actual robot state against the
expected target/trend. If feedback is missing or does not match, the run should
be diagnosed as retry/blocked instead of reported as physically complete.

The shipped skill surface is intentionally small and real-platform focused:

- `observe`
- `capture_image`
- `detect_grasps`
- `move_arm_ee`
- `move_arm_joints`
- `set_gripper`
- `set_base_velocity`
- `set_lift_height`
- `stop_motion`

`detect_grasps` wraps AnyGrasp SDK under `third_party/anygrasp_sdk`; runtime
requires a valid machine-bound license, a detection checkpoint, and AnyGrasp's
Python dependencies installed in the same uv environment. Run it with
`check_only=true` to get the exact missing package list. There are no complete
task policies in this tree. Learned skills can be added later under the user
skill root without mixing them into the platform adapter layer.

## Skill Self-Repair Policy

LoopMaster should not ask the operator to answer questions that can be resolved
from the repository. When a run fails because a shipped skill omits data, has a
schema mismatch, or has stale documentation, the agent loop should first inspect
the local skill implementation and trace evidence, update the relevant
`SKILL.md`/`policy.py` when it has write authority, and rerun the check. User
input is reserved for hardware safety decisions, missing task intent, external
runtime state, or changes that would create a new learned behavior.

For example, if `observe` succeeds but the trace only exposes `state_keys` while
downstream control needs numeric joint values, the correct repair is to update
the `observe` skill to return numeric `state` values in its observation summary,
then rerun. It is not a user-facing research question.

## HEI ReBot Lift Binding

The platform adapter wraps the HEI ReBot Lift LeRobot interfaces already present
at:

```text
hei-rebot-lift/software/lerobot-hei-rebot-lift/src/lerobot/robots/hei_rebot_lift/
```

Use `HeiRebotLiftPlatformConfig(remote_ip="...")` for the remote host/client
path, or omit `remote_ip` for direct local hardware access. The adapter imports
LeRobot lazily so framework tests and planning code do not require hardware
dependencies at import time.

## Smoke Run

```bash
uv sync
uv run python -m loopmaster_agentic "inspect robot state" --dry-run
```

For the full local robot/VR dependency set, use:

```bash
uv sync --all-extras --all-groups
```

Dry-run uses an in-memory platform only for framework verification. Production
use should pass a real `HeiRebotLiftPlatform` to `Handler.run(...)`.

## Persistent Handler Chat

Open a persistent terminal conversation with the Handler:

```bash
uv run python -m loopmaster_agentic chat --dry-run
```

Each non-command message is executed as one Handler run. `--dry-run` only swaps
the robot platform for `DryRunPlatform`; the Handler, Strategist, Worker, and
Auditor subagents still run through the `fnyweg` Codex profile by default. Use
`--agent-profile <name>` or `LOOPMASTER_CODEX_PROFILE=<name>` to select another
profile, and `--local-agents` only for offline/debug runs that should not call
Codex.

The transcript is saved under
`~/.loopmaster_agentic/handler_chat/handler-direct.jsonl` by default, and each
role's Codex session id is saved beside it, so the same `--session-id` resumes
prior turns across terminal sessions. Use `/history`, `/clear`, and `/exit`
inside the TUI.

For scripted checks without entering the TUI:

```bash
uv run python -m loopmaster_agentic chat --dry-run --once "inspect robot state"
```

loopmaster

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README