NVIDIA Automates Robot Navigation Training With COMPASS

NVIDIA has detailed a new agent-driven workflow that uses AI coding assistants like Codex or Claude Code to automate robot training. The system relies on the company's Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis, known as COMPASS. Instead of training a robot's navigation from scratch, COMPASS takes the pretrained NVIDIA X-Mobility base policy and uses reinforcement learning to train a 'residual specialist' that corrects actions for specific robots and environments.

The automated workflow guides developers through validation, scene preparation, smoke testing, residual training, checkpoint evaluation, and runtime packaging, using human approval gates to verify each step. To demonstrate the process, NVIDIA used the Boston Dynamics Spot quadruped across three scene paths. These include a built-in warehouse baseline, generated indoor environments from the SAGE-10K dataset—which contains 10,000 generated indoor scenes across 50 room types—and captured environments reconstructed using NVIDIA Omniverse NuRec.

To run the reference implementation, developers need NVIDIA Isaac Lab 3.0 and Isaac Sim 6.0. The hardware requirements include an Ubuntu 22.04 or 24.04 system with at least 32 GB of RAM, an RTX-capable GPU with at least 16 GB of VRAM, and Linux driver 580.95.05. The minimum reference GPU for Isaac Sim 6.0 is a GeForce RTX 4080. The setup also requires Docker Engine 24 or later with the NVIDIA Container Toolkit, alongside a Hugging Face account to download the gated COMPASS assets and X-Mobility checkpoints.

At runtime, the exported policy operates without the AI agent. It processes RGB camera inputs, odometry, and a goal point to publish velocity commands on the /cmd_vel topic. If a robot lacks compatible state estimation, developers can deploy NVIDIA cuVSLAM, a CUDA-accelerated visual library, to provide visual odometry mapped to the /chassis/odom topic. Performance is measured using standard COMPASS metrics, including goal-reached rate, fall-down rate, and travel time.