NVIDIA Unveils COMPASS: An AI-Driven Framework for Simplifying Robot Navigation Training

NVIDIA has introduced COMPASS (Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis), an AI-driven framework designed to simplify robot navigation training. This new tool leverages reinforcement learning (RL) and pre-trained navigation models to reduce the time and cost of adapting navigation capabilities across different robot types and environments.

The core idea behind COMPASS is to avoid reinventing the wheel for every robot-environment combination, a traditional method that requires rebuilding datasets, simulations, and training pipelines. Instead, COMPASS uses a pre-trained base policy, NVIDIA's X-Mobility model, and fine-tunes it with residual RL to adapt navigation actions for specific robots and environments.

The framework automates routine tasks such as dependency validation, simulation asset preparation, and training diagnostics, integrating human approval checkpoints at key stages. Developers can test the system using three primary options: a preconfigured warehouse simulation, a generated scene from the SAGE-10K dataset, or a real-world environment reconstructed using NVIDIA's Omniverse NuRec.