COMPASS Packages Cross-Robot Navigation Training as Codex/Claude Code Skills

NVIDIA has released an agentic training workflow for COMPASS, packaging operations previously spread across Isaac Sim, Isaac Lab, scene conversion, reinforcement learning, and ROS 2 as repository skills. Developers can use `$compass` in Codex or invoke `/compass` in Claude Code to ask a coding agent to check container, GPU, model, and asset versions; create an occupancy map; run a single-environment smoke test; launch long-running training; and save logs, telemetry, checkpoints, and an artifact manifest.

COMPASS does not allow a language model to control a robot directly. Under the hood, it first reuses X-Mobility’s imitation learning policy, then applies residual reinforcement learning to train a corrective policy for a specific robot and scene. Multiple specialists can subsequently be distilled into a shared cross-embodiment policy. The official model card lists G1, H1, Carter, and Spot. Inputs include RGB images, robot velocity, a path, and embodiment type, while outputs are linear and angular velocity. The reference ROS 2 integration converts camera data, odometry, and goals into model inputs, then publishes control commands to `/cmd_vel`.

The new workflow supports three scene paths: a registered warehouse can serve as a reproducible baseline; SAGE-10K provides 10,000 generated indoor scenes spanning 50 room types; and NuRec converts real-world stereo RGB captures into reconstructed environments usable in Isaac Sim. The agent must pause for human approval before scene acceptance, smoke testing, and checkpoint promotion. Evaluation also requires the base policy and residual checkpoint to use the same seed, goals, initial states, and termination conditions, while comparing success rate, fall rate, and travel time.

The technical value of this design lies in turning a robot-training pipeline—one that can easily fail because of versions, coordinate systems, or scene assets—into an auditable agentic workflow, rather than claiming a sudden improvement in model capabilities. NVIDIA has not published success rates, training times, or cross-robot gains resulting from the introduction of these Skills. The tutorial also stops at checkpoint evaluation; ONNX/TensorRT export and deployment on physical robots still require separate validation. The code is licensed under Apache 2.0, but the models and simulation assets are subject to NVIDIA’s terms. NVIDIA also explicitly states that COMPASS has not been validated for functional safety or highly dynamic environments.