Skip to content

srirangam-r/robot-harness

v0.1.0

Drive a simulated robot arm (LIBERO / robosuite) through three tools: observe, run a skill program, done. Motions come from recorded human demos (motor memory).

GitHub

robot-harness power

Packages the Motor Memory robot MCP server (observe / run / done) and a skill that teaches an agent the robot protocol (turn format, skill grammar, status words, retry rules). Agent Plugins 1.0.0 layout:

plugin.json                       manifest
mcp.json                          robot MCP server: python -m motormem.brain.mcp_server
skills/robot-harness/SKILL.md     the robot protocol

Requirements

  • python on PATH must import motormem and mcp: start Kiro from the activated repo venv (or pip install -e .). Otherwise the robot server connects with no tools; set command in mcp.json to the venv's absolute python path.
  • Real simulator backend: LIBERO/robosuite installed, a GPU with EGL, and a built index (data/index/*.pkl).

Install

kiro-cli powers install ./powers/robot-harness      # from the repo root
# or, inside a Kiro chat:  /powers install /abs/path/to/powers/robot-harness
kiro-cli powers uninstall robot-harness

Configure an episode

The server reads its episode from ROBOT_EPISODE, which defaults to ${PLUGIN_DATA}/episode.json (Kiro's per-power data directory). Without that file every tool answers ERROR: no episode configured. Write a spec there, e.g. the toy backend (no simulator needed):

{"suite": "libero_goal", "task_id": 8, "init_index": 0, "out_dir": "/tmp/robot_ep", "max_steps": 600}

For the real simulator change MOTORMEM_BACKEND in mcp.json to real (it overrides the spec's backend key) and add "index_path" to the spec. Add "record": true to save video frames to out_dir, then render with python scripts/make_video.py <out_dir>.

Fairness: the server strips privileged simulator state from every observation and never reports task success to the agent (see .kiro/steering/fairness.md).