This article is a technical commentary and implementation example created using AI. Although the code and procedures presented are based on primary sources, the author has not verified their operation on actual hardware. Operation may vary depending on the environment and version.
Training Cross-Embodiment Robot Navigation Policies and the COMPASS Framework Explained via Official Information
Robot navigation features are essential for understanding the surrounding situation, selecting paths, avoiding obstacles, and safely reaching target destinations. However, every time this functionality is migrated to a different robot or environment, the effort required to repeat new data, simulation assets, robot interfaces, training, diagnostics, and evaluation presents a significant challenge.
Primary information published on the NVIDIA Technical Blog explains “COMPASS,” a framework that adapts pre-trained “NVIDIA X-Mobility” policies into residual specialists tailored for specific robots and environments using a workflow combining AI agents and human-in-the-loop approval gates. Based on official information, this article organizes its components and procedures.
What is COMPASS
COMPASS (Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis) is an integrated framework that achieves scalable cross-embodiment mobility by leveraging expert demonstrations from a single embodiment.
Rather than learning navigation behaviors entirely from scratch, it reuses the behavior of pre-trained “NVIDIA X-Mobility” policies and trains “residual specialists,” which are reinforcement learning (RL) policies that correct the base behavior to match the selected robot and environment. Data from multiple specialists can then be distilled into a shared cross-embodiment policy.
flowchart TD
A["NVIDIA X-Mobility Base Policy"] --> B["COMPASS Framework"]
B --> C["Embodiment Specialists via Residual RL"]
C --> D["Shared Cross-Embodiment Policy"]
Overview of the Agent-Driven Workflow
Developers define the robot, scene sources, and navigation goals, while coding agents use repository skills to validate dependencies, prepare assets, run smoke tests, launch training, diagnose failures, and compare checkpoints. Human approval gates retain control at critical milestones, such as scene acceptance, smoke testing in a single environment, and checkpoint promotion.
The reference implementation adopts Boston Dynamics Spot, a quadruped robot, and provides three paths: a built-in warehouse, generated SAGE-10K indoor scenes, and environments reconstructed using NVIDIA Omniverse NuRec.
Additionally, if compatible odometry or transforms are not provided on the robot side, “NVIDIA cuVSLAM,” a CUDA-accelerated visual odometry and SLAM library, can be utilized as deployment odometry.
Development Environment Requirements
The primary information lists the following hardware and software requirements for utilizing the COMPASS software stack:
Ubuntu 22.04 or 24.04 system
32 GB or more RAM
RTX-capable NVIDIA GPU with 16 GB or more VRAM (Isaac Sim 6.0 minimum reference GPU is GeForce RTX 4080, tested Linux driver is 580.95.05)
Docker Engine 24 or later with NVIDIA Container Toolkit
Hugging Face account with access to gated nvidia/COMPASS and nvidia/X-Mobility Hugging Face repositories and a read token
Tested stack: NVIDIA Isaac Lab 3.0 and NVIDIA Isaac Sim 6.0
Step 1: Setting Up the COMPASS Agent Workflow
Prepare the repository and instruct the coding agent on the workflow contract before starting scene work. Download gated assets, enable COMPASS skills in Codex, validate the stack, and pause at the approval gate for a single environment.
For example, by exposing skills in a Codex or Claude Code environment and passing validation prompts, you generate reviewable evidence at each stage, such as software and asset inventories, environmental reports, and smoke test logs. Sensitive information such as tokens must be handled securely and configured only on the shell.
Step 2: Selecting and Preparing Navigation Scenes
You can choose from the following three paths as scene sources:
Built-in Warehouse (combined_multi_rack): The fastest reproducible baseline. Robots, scenes, and occupancy maps are already registered, making it optimal for installation verification upon deployment.
SAGE-10K Scenes: Contains 10,000 generated indoor scenes spanning 50 room types, providing geometry, materials, layout metadata, and previews. Used after passing approval gates such as USD confirmation and occupancy map validation in Isaac Sim.
NuRec Captured Environment: An optional path for converting stereo RGB captures into Isaac Sim-compatible reconstructions for simulation evaluation and fine-tuning in deployment environments.
Step 3: Integrated Validation of Robot and Scene
Once the selected scene is available in COMPASS, run a preview in a single environment before scaling up training. Verify that Isaac Sim launches, the scene loads, the robot spawns in a valid position, camera observations are acquired, and it responds to policy commands without clipping, tipping, or unresolved simulation errors.
Step 4: Training Residual Specialists
After preview approval, launch the standard residual RL workflow.
[To be verified in Windows environment]
python run.py -c configs/train_config.gin -o ./outputs/spot_combined_multi_rack -b ./assets/x_mobility.ckpt --enable_cameras --embodiment spot --environment combined_multi_rack
During training, set the number of environments according to GPU memory, and monitor reward components, goal progress, contacts/falls, episode termination, throughput, and GPU memory. Do not assume the final iteration is optimal; periodically save checkpoints and evaluate under matched conditions. If failures occur, investigate using the diagnostic workflow before changing environments or settings.
Conclusion
By leveraging the COMPASS framework, you can perform efficient residual learning and scene validation while building upon pre-trained NVIDIA X-Mobility policies, combining AI agent assistance and human approval gates. It is important to refer to the official repository and guides and proceed with implementation while verifying procedures in an appropriate environment.
コメント