A step-by-step tutorial shows how to use coding agents like Codex or Claude Code to run the COMPASS (Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis) workflow, which trains a residual reinforcement learning policy on top of NVIDIA's pretrained X-Mobility navigation model. Using the Boston Dynamics Spot robot as a reference, the workflow covers repository setup, scene selection (built-in warehouse, SAGE-10K generated scenes, or Omniverse NuRec captured environments), smoke testing, residual training, checkpoint evaluation, and runtime integration with optional cuVSLAM odometry. Human approval gates control scene acceptance, smoke tests, and checkpoint promotion throughout the agent-driven process.
Table of contents
What is COMPASS?Step 1: Set up the COMPASS agentic workflowStep 2: Choose and prepare a navigation sceneStep 3: Validate the robot-scene integrationStep 4: Train the residual specialistStep 5: Evaluate before promoting a checkpointStep 6: Connect the policy to the robot runtimeGet started with COMPASSQuestions this post answers
What is COMPASS in the context of robot navigation policies?
COMPASS is a unified framework for scalable cross-embodiment mobility that reuses navigation behavior from a pretrained X-Mobility policy and trains a residual reinforcement learning specialist to correct the base action for a specific robot and environment, rather than relearning navigation from scratch. Data from multiple specialists can later be distilled into a shared cross-embodiment policy. Robotics engineers evaluating agentic training workflows can follow ongoing COMPASS coverage on daily.dev.
What hardware do I need to run Isaac Sim 6.0 for COMPASS robot navigation training?
An Ubuntu 22.04 or 24.04 system needs at least 32 GB of RAM, an RTX-capable NVIDIA GPU with at least 16 GB of VRAM, and Linux driver 580.95.05, the version tested for Isaac Sim 6.0. The minimum reference GPU is a GeForce RTX 4080, and Docker Engine 24 or later with the NVIDIA Container Toolkit is also required. Engineers planning simulation infrastructure can track hardware requirement changes for tools like Isaac Sim on daily.dev.
What metrics does COMPASS use to evaluate a residual training checkpoint before promotion?
Standard COMPASS evaluation reports goal-reached rate, fall-down rate, and travel time when comparing the pretrained X-Mobility base policy against residual checkpoints under matched seeds, goals, initial states, rollout length, and active terminations. Additional evidence like goal progress, contact behavior, or timeouts must be labeled separately as derived or custom instrumentation, and a human must approve promotion. Teams building evaluation gates into ML pipelines can compare approaches like this on daily.dev.