Unitree Hermes Colab

Goal

Create or review a Colab notebook that makes Hermes Agent useful for Unitree

work without pretending Colab is a safe robot-control host. The notebook should

install or check Hermes, clone selected Unitree repositories, generate

read-only artifacts, and show clear pass/fail review gates.

Hard Boundaries

• Do not execute robot-control commands from Colab.
• Do not publish DDS, ROS, motor, sport-mode, or low-level commands.
• Do not SSH, SCP, tunnel, or scan Unitree robot LAN addresses such as

192.168.123.0/24.

• Do not claim physical hardware validation unless the user provided external

evidence.

• Put risky local-host commands in quoted runbooks for a human to review and

run on the correct machine.

• Hermes one-shot execution must be opt-in. It is acceptable to install Hermes

and prepare prompts by default.

Build Workflow

1. Keep one configuration block near the top of the notebook or runner:

INSTALL_HERMES, RUN_HERMES_AGENT, CLONE_UNITREE_REPOS, PROVIDER,

MODEL, and output directory.

1. Record runtime versions: Python, platform, GPU/CUDA when present, Hermes CLI

status, and cloned repo commit SHAs.

1. Clone only important Unitree repositories by default:

unitreerobotics/xr_teleoperate, unitree_sdk2_python, and

unitree_mujoco.

1. Write a local AGENTS.md safety file before any optional Hermes run.
2. Generate these use cases at minimum:

simulation runbook, teleoperation preflight checklist, log triage,

contribution scouting, and IK evidence review.

1. Render user-facing outputs as a report: cards, repository map, flow

visualization, review gates, references, and saved artifacts.

1. Save unitree-hermes-report.md, unitree-hermes-review.json, AGENTS.md,

and a flow visualization.

Output Standards

• Do not leak meta-instructions like "Notebook organization" into the report.

Use user-facing labels such as "Review gates", "Runtime", "Use cases", and

"References".

• Keep references limited to important sources: Hermes Agent, Hermes docs,

Unitree repositories, and Codex skills/subagents when the artifact includes a

Codex skill.

• Include a critical usefulness score. Good default framing:

high value for setup review and log triage, low value for live robot control.

• Mark skipped Hermes execution clearly when no provider key is present or

RUN_HERMES_AGENT is false.

Validation

Run the project checks when working in this repo:

python3 -m py_compile src/unitree_colab_ik/hermes_lab.py src/unitree_colab_ik/hermes_cli.py notebooks/run_unitree_hermes_agent_lab.py
python3 -m json.tool notebooks/unitree_hermes_agent_lab.ipynb >/dev/null
python3 -m compileall -q src tests
python3 skill/unitree-hermes-colab/scripts/check_lab_artifacts.py <path-to-unitree-hermes-review.json>

Run tests if pytest is available:

python3 -m pytest tests/test_hermes_lab.py

Subagent Review

Use Codex subagents only when the user explicitly asks for parallel/subagent

review. Keep them read-heavy. Good split:

• safety reviewer: checks no robot-control execution path exists
• notebook reviewer: checks clean-runtime reproducibility and visible outputs
• contribution reviewer: checks whether the Unitree/Hermes use cases are

genuinely useful and small enough to publish

The main agent should wait for summaries and integrate findings. Avoid

parallel write-heavy edits unless the user explicitly requests that.