LoopMaster Agentic Robotics

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:

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

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

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

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:

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:

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