NVIDIA COMPASS Uses Coding Agents to Adapt Robot Navigation Policies

NVIDIA published an agent-driven workflow for cross-embodiment robot navigation

NVIDIA published a new COMPASS development workflow on August 26, 2026 for adapting robot navigation policies across different embodiments and environments.

COMPASS stands for Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis. The framework starts from NVIDIA's pretrained X-Mobility navigation policy, then trains residual reinforcement-learning specialists that correct the base action for a particular robot and scene instead of relearning navigation from scratch.

The new developer workflow packages much of that process into repository skills that coding agents such as Codex or Claude Code can invoke during setup, scene preparation, smoke testing, training, diagnosis and checkpoint evaluation.

The coding agent does not control the physical robot at runtime. NVIDIA explicitly positions it as a development-time orchestrator around a human-reviewed training and validation process.

Residual RL adapts a base policy instead of starting over

Cross-embodiment robotics is difficult because a policy trained for one platform may not transfer cleanly to another body, sensor layout or control interface.

COMPASS addresses that by reusing the X-Mobility base policy and learning a residual specialist for the target robot-scene combination.

That residual policy learns corrections to the base action. Multiple specialists can later contribute data to a shared cross-embodiment policy.

The approach is intended to reduce the cost of collecting demonstrations and retraining navigation behavior independently for every new robot and environment.

Coding-agent skills automate the development workflow, with stop points for humans

The new workflow exposes skills such as $compass, $compass-doctor and $compass-newembodiment.

A developer can ask the coding agent to validate dependencies, prepare simulation assets, run a one-environment smoke test, launch residual training, inspect logs and compare checkpoints.

NVIDIA repeatedly places human approval gates before consequential transitions. The workflow pauses for scene acceptance, smoke-test approval and checkpoint promotion.

That is an important design choice. A coding agent can automate repetitive setup and evaluation tasks, but the framework does not assume that an autonomous agent should silently change dependencies, rewards, scene assets or promote a robot policy without review.

The reference path uses Boston Dynamics Spot

NVIDIA's tutorial uses the Boston Dynamics Spot quadruped as the reference robot.

Developers can start with a registered warehouse scene, move to generated SAGE-10K indoor scenes, or use NVIDIA Omniverse NuRec to reconstruct a captured physical environment for simulation.

For generated or reconstructed scenes, the workflow includes checks for geometry, scale, collision meshes, occupancy maps, robot clearance and a one-environment preview before full training.

This is especially relevant in robotics because a visually plausible scene can still be physically wrong. Bad collision geometry or coordinate conventions can produce misleading training results.

Training produces residual checkpoints that must be evaluated under matched conditions

After a successful smoke test, the workflow launches residual RL in NVIDIA Isaac Lab and Isaac Sim.

NVIDIA recommends saving periodic checkpoints rather than assuming the last checkpoint is best.

Evaluation compares the pretrained X-Mobility base policy with residual candidates under matched seeds, goals, initial states, rollout lengths and termination settings.

Standard COMPASS metrics include goal-reached rate, fall-down rate and travel time.

Any extra metrics such as goal progress, contact behavior or command stability should be clearly labeled as derived or custom evidence.

The workflow then stops for human approval before packaging or promoting a checkpoint.

ROS 2 connects the trained policy to robot runtime

The trained navigation policy receives front-camera imagery, odometry, a goal point and optional map/route context.

The reference integration publishes forward-linear and angular velocity commands through ROS 2 on /cmd_vel.

NVIDIA says optional cuVSLAM can provide camera-based odometry when the target robot does not already expose compatible state estimation.

The company is careful to separate this runtime architecture from the coding agent. The agent helps create and validate the policy during development; the exported policy and robot controller perform navigation at runtime.

New embodiments require a separate onboarding workflow

For robots not already registered in COMPASS, NVIDIA provides a $compass-newembodiment path.

That process covers robot configuration, scene registration, action mapping and visual smoke testing before specialist training begins.

Export to ONNX, JIT or TensorRT and deployment to physical hardware also require separate target-specific validation.

NVIDIA does not define a universal success threshold because training time, safety behavior and checkpoint quality depend on the robot, environment, reward design and compute configuration.

Why this matters

The important development is not simply "AI agents control robots." That would be inaccurate.

Instead, NVIDIA is using coding agents as workflow operators for robot-learning engineering: they validate environments, prepare assets, start training, diagnose failures and assemble comparable evidence while humans retain approval over critical gates.

That pattern can make robotics experiments more reproducible because the agent is asked to preserve exact commands, repository revisions, configuration files, logs, videos, checkpoints and evaluation records.

It also shows a broader direction for agentic development tooling. As SDKs become more complex, vendors are packaging operational knowledge as agent skills so developers can invoke versioned procedures instead of manually reconstructing long setup guides.

The value will depend on how well those skills behave across real hardware, custom robots and failure cases, but COMPASS provides a concrete example of human-supervised agentic automation inside a serious robotics training loop.