---

name: robot-deployment description: "Prepare and review Unitree Go1 deployment artifacts, LCM contracts, RC command and gait profiles, calibration and emergency gates, logging, and network/container prerequisites without actuating a robot or mutating a deployment host." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Robot Deployment

Use this sub-skill for deployment preparation and review only. It distills the repository's Go1 low-level controller boundary: the policy/checkpoint contract, LCM observations and targets, RC commands and gait profiles, calibration, emergency behavior, logging, and the Jetson/container/network prerequisites.

Safety gate — read before every other step

1. Do not actuate. This skill never publishes LCM targets, starts the Unitree SDK, launches a controller, loads a Docker image, transfers files, changes a route/interface, or runs the architecture-specific lcm_position binary. It has no bundled launcher.
2. Real end-to-end deployment is BLOCKED_REQUIRED_BACKEND unless a human operator supplies a Unitree Go1 Edu, a compatible onboard computer or approved external controller, the matching SDK/LCM runtime, a physically controlled test area, and explicit approval for each unsafe step. Static artifact, schema, and read-only diagnosis work can continue without hardware.
3. Before any operator-run activity, use Safety: suspend the robot, clear the area, enter damping mode, keep people/obstacles away, do not attach Ethernet/USB at the rear during initial calibration or a new-policy test, and confirm the operator has an emergency stop plan. The README marks this research code as use-at-your-own-risk.
4. Every transfer, SDK start, container start, calibration, and control-loop launch is human-approved and unsafe. This skill provides an order and gates, not copy-paste execution. Stop on any missing prerequisite or unexpected motion.

Route and operating workflow

1. Read Safety, then Controller workflow.
2. Route policy labels, parameters.pkl, observation-history shape, TorchScript export, and checkpoint repair to training-and-policy; route actuator fitting and deployment-log sample preparation to actuator-network.
3. Validate a candidate policy directory with the read-only scripts/validate_policy_artifacts.py. It checks names and metadata only; it never deserializes a pickle, loads TorchScript, runs inference, or enters a control loop.
4. Check LCM contract, then use scripts/diagnose_network.py for a read-only local interface/IP/multicast report. A report is not proof that a robot is reachable and does not modify the host.
5. Review Installation and transfer and Container runtime. The operator must independently approve credentials, disk, architecture, image provenance, network configuration, and rollback before using SSH/rsync/Docker/SDK tools.
6. Use Troubleshooting for failures. If a required hardware/backend, generated message module, matching binary, or checkpoint component is unavailable, report the blocker rather than trying a side effect.

Contracts worth preserving

• The deployment script constructs an LCM agent and RC profile at 0.02 s (50 Hz) and wraps the observation in a history buffer before calling the exported policy. The effective agent step is control.decimation * sim.dt; it must agree with the policy's training contract.
• The policy's first twelve outputs are joint actions. The deployment agent applies action scale, hip-scale reduction, default joint pose, and the source permutation before forming a pd_tau_targets message. Do not hand-edit this mapping; see the controller and LCM references.
• State input is assembled from gravity, scaled commands, joint position/velocity deltas, prior actions, optional velocity/yaw/contact/clock/height fields, and the configured observation history. A configuration mismatch is a hard stop, not an invitation to guess.
• udpm://239.255.76.67:7667?ttl=255 is the source LCM URL. State subscriptions and the output target channel are documented, but this skill never subscribes or publishes at runtime.

Safe validation outputs

A useful handoff consists of: the exact candidate artifact directory; the validator's pass/fail report; the policy config and expected observation/action shapes; the planned SDK → controller order; LCM URL/interface evidence; human safety approval; and rollback/stop criteria. No private environment paths, credentials, copied model weights, generated logs, or vendor binaries belong in this skill.

References

• safety.md — physical, damping, calibration, PowerProtect, and emergency gates.
• controller-workflow.md — transfer/controller order, policy contract, mapping, timing, RC/gaits, calibration, and logging.
• lcm-contract.md — URL, channels, schemas, state fields, camera caveat, and architecture boundary.
• installation-and-transfer.md — human-run SSH/rsync/Docker sequence, prerequisites, checkpoints, and rollback.
• container-runtime.md — Jetson L4T, privileged host-network NVIDIA Docker, ARM, and display assumptions.
• troubleshooting.md — known architecture, network, dependency, storage, checkpoint, tilt, and camera failures.

Evidence boundary

This operating context was distilled from the repository README deployment and debug sections, go1_gym_deploy/scripts/deploy_policy.py, utils/deployment_runner.py, utils/command_profile.py, utils/cheetah_state_estimator.py, utils/network_config_unitree.py, utils/logger.py, envs/lcm_agent.py, envs/history_wrapper.py, generated LCM schemas, and Docker/autostart/installer/transfer files. The shell scripts are evidence of side effects, not runtime helpers. No claim here substitutes for Unitree safety documentation or a hardware-qualified operator.