NVIDIA COMPASS

NVIDIA COMPASS is a framework and trained model for cross-embodiment robot navigation. Its name expands to Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis. COMPASS starts from NVIDIA's X-Mobility navigation policy, adapts that policy to individual robot bodies with residual reinforcement learning, and distills the resulting specialists into one policy conditioned on robot identity.^[1][2]

The system is designed for point-to-point mobility rather than joint-level locomotion. It maps camera observations, measured motion, route or goal information, and an embodiment identifier to desired linear and angular velocity. A separate low-level controller turns those velocity commands into wheel or joint actuation.^[1][3] This page concerns the NVIDIA robotics project, not Compass, the Shopee language model.

An arXiv preprint was first submitted in February 2025 and revised in October 2025. NVIDIA later listed COMPASS among its accepted papers for ICRA 2026.^[1][4] The implementation and training code are public, while the pretrained checkpoint and simulation assets require acceptance of NVIDIA's gated model terms.^[2][3]

Design

COMPASS separates reusable navigation knowledge from adaptation to a particular robot. The published workflow has three stages.^[1]

Scroll sideways for more →

X-Mobility base policy

X-Mobility combines an autoregressive world model with a policy head. The world model encodes recent RGB observations, measured velocity, and previous actions into a latent state. A route encoder adds goal guidance, and the policy predicts a velocity command. COMPASS freezes an X-Mobility checkpoint trained on Nova Carter and reuses that representation for subsequent robot-specific training.^[1][5]

The base model is an imitation learning policy, so it can inherit errors associated with the teacher data and with closed-loop states that were rare or absent in those demonstrations. The COMPASS authors use reinforcement learning in simulation to expose a robot to its own action consequences and learn corrective behavior.^[1]

Residual specialists

For each robot, COMPASS adds the base action and a learned residual action. The base policy stays frozen while the residual network is optimized with Proximal Policy Optimization. Its reward includes progress toward the goal, penalties for collision or falling, and a completion reward for reaching the destination at zero velocity.^[1]

This design does not remove platform integration work. Humanoids and quadrupeds in the paper use locomotion policies that translate COMPASS velocity commands into joint controls. Nova Carter uses a custom simulated root-state controller because of the wheel-physics limitations reported by the authors. Adding an unregistered robot also requires an articulation configuration, environment registration, an action mapping, and a smoke test before specialist training begins.^[1][6]

Generalist distillation

The specialist rollouts record the latent policy state, an embodiment identifier, and the mean and variance of the specialist's Gaussian action distribution. The generalist is trained to minimize Kullback-Leibler divergence from these recorded specialist distributions. In the published experiments, the identifier is a one-hot vector with one entry per robot.^[1]

The phrase "cross-embodiment" therefore needs a boundary. The H1, G1, and Spot platforms were unseen by the original Carter-only imitation-learning stage, but each received residual training and an identity entry before distillation. The paper leaves learned embeddings for zero-shot generalization to a wholly new robot body as future work. COMPASS does not establish that an arbitrary new robot can use the released generalist without registration, action-interface engineering, or adaptation.^[1][6]

Training implementation

The paper trains specialists in NVIDIA Isaac Lab across Nova Carter, Unitree H1, Unitree G1, and a Spot quadruped. Goals are sampled two to five metres from the initial pose, and an episode ends after a collision, successful arrival, or 256 simulation steps. Each specialist was trained with 64 parallel environments for 1,000 episodes on two NVIDIA L40 GPUs, except Carter, which used 300 episodes because the base checkpoint was already trained on that platform.^[1]

For distillation, the authors recorded 320 trajectories of 128 steps for each embodiment, approximately 40,000 frames per robot, then trained on four NVIDIA H100 GPUs.^[1] The public model card describes a vision-transformer and multilayer-perceptron architecture and names G1, H1, Carter, and Spot as the enabled embodiment identifiers for model version 1.0.^[3]

The software repository has continued beyond the paper configuration. The handbook lists Carter, H1, G1, Spot, and Digit as configurable embodiments, while the released model card names only the first four. A code path or robot configuration is not, by itself, evidence that the released checkpoint was evaluated on that platform.^[3][7]

Reported evaluation

The paper evaluated X-Mobility, each residual specialist, and the distilled generalist in four simulated layouts. Each layout used 640 trials per robot with randomized initial and goal poses and a 25.6-second timeout. Success meant reaching the goal without a collision or timeout.^[1]

The success rates below are author-reported. They have not been converted into claims about performance outside the paper's simulator, scenes, routes, or termination rules.^[1]

Scroll sideways for more →

For H1, Spot, and G1, which were unseen during Carter-only imitation learning, the paper summarizes residual training as raising success by approximately 5 to 40 times over the base policy and reducing the weighted travel-time metric by about threefold on average. This is the authors' aggregate description; the table's cell-by-cell ratios vary by robot and scene. The distilled policy stayed close to the corresponding specialist in most reported cells. An RL-from-scratch comparison using the same latent state did not converge within the authors' 1,000-episode run.^[1]

For physical deployment, the authors ran the generalist on Carter and G1 using NVIDIA Jetson Orin computers and cuVSLAM localization. They report approximately 80 percent success across 20 trials with randomly placed obstacles and no additional real-world fine-tuning. The small trial count, two selected platforms, and author-run protocol limit what can be inferred about broader reliability.^[1]

The main simulation and physical-deployment results above come from the COMPASS authors and NVIDIA.^{[1][2][3][6][8]}

Independent evaluation

Rahmi Kavak and Ozgur Yilmaz independently evaluated a trained COMPASS policy in static warehouse layouts and variants with moving human-like agents. Their ICHORA 2026 paper's published abstract reports success of up to 92.0 percent and Base Contact Rate of about 8 to 11 percent in static settings, compared with about 84 to 85 percent success and about 15 percent Base Contact Rate under added motion. The authors present this as a baseline for socially aware navigation and report reduced interaction safety under motion. It is not a replication of NVIDIA's four-embodiment protocol.^[11]

Agent-assisted development workflow

In August 2026, NVIDIA documented a workflow in which a coding agent helps developers validate dependencies, prepare a scene, run a one-environment smoke test, launch residual training, diagnose failures, compare checkpoints, and package an approved result. The reference tutorial uses Spot in a built-in combined_multi_rack warehouse, a generated SAGE-10K indoor scene, or an optional scene reconstructed with Omniverse NuRec.^[8]

The repository exposes three related skills. compass orchestrates training, evaluation, scene preparation, and deployment work; compass-doctor performs read-only diagnostics; and compass-newembodiment guides the configuration of a new robot. The workflow places human approval gates after scene inspection, the one-environment smoke test, and checkpoint evaluation.^[6][8]

The coding agent is a development tool, not part of the deployed control loop. After training and export, the COMPASS policy and the robot's controller perform navigation without the coding agent. NVIDIA's tutorial explicitly separates agent-assisted development from runtime control.^[8]

Release 1.6.0, published in May 2026, added OSMO workflow submission, distributed multi-GPU training, automatic occupancy-map generation, container-based development, the agentic skills, and a no-regression benchmark gate. The tagged release updated the repository to Isaac Lab 3.0.0-beta1. NVIDIA's later August tutorial states that its tested setup used Isaac Lab 3.0 with NVIDIA Isaac Sim 6.0; developers need to follow the versions pinned by the repository and tutorial they are reproducing.^[9][8]

Runtime and deployment

The repository can export policies to ONNX, TorchScript, and TensorRT. Its reference Robot Operating System 2 integration turns a front-camera image, navigation target or route, and robot speed derived from odometry into policy inputs. The inference node publishes forward-linear and angular-velocity commands on /cmd_vel; recurrent state and the previous action remain internal to the inference implementation.^[6][8][10]

cuVSLAM can supply camera-based odometry when a robot lacks compatible state estimation. It runs as a separate ROS 2 component and is not part of COMPASS training. Its map is also not an input to the navigation policy, even though its odometry can support the navigator.^[8]

The model card reports TensorRT inference on Jetson Orin and an output vector of six desired linear and angular velocity values. The deployment still requires validation of coordinate frames, normalization, command limits, update rates, stopping behavior, and the target robot's low-level controller.^[3][8]

Availability and limitations

NVIDIA publishes the COMPASS implementation under the Apache License 2.0.^[2] The nvidia/COMPASS repository on Hugging Face is gated, and its checkpoint and packaged assets use the NVIDIA Open Models License rather than the code license. The model repository includes ONNX and TorchScript artifacts for Carter, simulation assets, and a G1 dataset for post-training GR00T.^[3]

The model card says COMPASS has not been tested or intended for mission-critical uses requiring functional safety. It identifies highly dynamic environments, limited embodiment support, and integration into customer environments as risks.^[3] The paper also reports lower success in office and long-horizon multi-rack tasks, proposes a hierarchical route planner as future work, and notes that one set of hyperparameters can diverge across robot bodies.^[1]

These limitations make the framework a research and engineering starting point rather than a certified navigation product. A deployment needs platform-specific safety controls, matched-condition evaluation, and tests in the intended environment. NVIDIA's agent-assisted tutorial deliberately avoids a universal checkpoint threshold and requires a human decision before a trained checkpoint is promoted.^[8]

References

1. ^Wei Liu et al. "COMPASS: Cross-embodiment Mobility Policy via Residual RL and Skill Synthesis." arXiv:2502.16372v3, October 27, 2025. arxiv.org/...2502.16372v3
2. ^NVIDIA Research. "COMPASS: Cross-Embodiment Mobility Policy via Residual RL and Skill Synthesis." Official repository. Accessed August 28, 2026. github.com/...COMPASS
3. ^NVIDIA. "nvidia/COMPASS." Model card and gated model repository. Accessed August 28, 2026. huggingface.co/...COMPASS
4. ^NVIDIA. "NVIDIA at ICRA 2026." Accessed August 28, 2026. nvidia.com/...icra
5. ^Wei Liu et al. "X-Mobility: End-to-End Generalizable Navigation via World Modeling." Proceedings of the 2025 IEEE International Conference on Robotics and Automation, pages 7569-7576. doi.org/...ICRA55743.2025.11128692
6. ^NVIDIA Research. "Agentic skills." COMPASS Handbook. Accessed August 28, 2026. nvlabs.github.io/...agentic
7. ^NVIDIA Research. "Quick start." COMPASS Handbook. Accessed August 28, 2026. nvlabs.github.io/...quickstart
8. ^Yan Chang, Mihir Acharya, Wei Liu, Katie Washabaugh, and Aishwarya Singh. "How to Train a Cross-Embodiment Robot Navigation Policy with AI Agents." NVIDIA Technical Blog, August 26, 2026. developer.nvidia.com/...tion-policy-with-ai-agents
9. ^NVIDIA Research. "COMPASS v1.6.0." GitHub release, May 18, 2026. github.com/...v1.6.0
10. ^NVIDIA Research. "ROS2 Deployment." COMPASS Handbook. Accessed August 28, 2026. nvlabs.github.io/...ros2
11. ^Rahmi Kavak and Ozgur Yilmaz. "Evaluating COMPASS for Socially-Aware Robot Navigation in Dynamic Indoor Environments." 2026 8th International Congress on Human-Computer Interaction, Optimization and Robotic Applications (ICHORA), pages 1-6, May 21, 2026. doi.org/...ICHORA69329.2026.11537134