COMPASS

NVIDIA·UC Berkeley·UT Austin · 2025-02

VANavigationNot disclosedRL policy

Trains per-embodiment residual RL experts on top of an imitation-learning-based policy, then distills them into a single embodiment-conditioned policy to unify wheeled·quadruped·humanoid navigation.

View architecture diagram in the gallery Paper ↗

Paper

COMPASS: Cross-embodiment Mobility Policy via Residual RL and Skill SynthesisarXiv 2502.16372

Overview

Category: Navigation
Type: VA
Released: 2025-02
Organization: NVIDIA·UC Berkeley·UT Austin
Open weights: Partial
Parameters: Not disclosed

Architecture

Backbone LLM: None
Vision encoder: DINOv2 (X-Mobility encoder)
Action head: Gaussian MLP policy (mean + global variance) → velocity commands (RL policy)
Chunk: None (1 step)
Control rate: Not disclosed (Jetson P50 29.3 ms)
Architecture (original summary): Policy based on frozen X-Mobility (DINOv2·latent world model) + per-embodiment residual RL (PPO) experts → distilled into an MLP generalist conditioned on one-hot embodiment embedding (mean prediction + global variance); outputs velocity commands (v, ω), legged robots use RL locomotion policy for joint control

Training

Pretraining accelerators: N/A (reuses the X-Mobility pretrained checkpoint)
Pretraining duration: N/A
Pretraining steps & batch: N/A
GPU/TPU-hours: Not disclosed
Fine-tuning: Residual RL: 2× L40, 64 parallel environments, 1,000 episodes per expert (Carter 300); distillation: 4× H100; time not disclosed
Data type: Simulation·synthetic
Data scale: Distillation data: 320 trajectories × 128 steps per embodiment ≈ 40k frames (HF card: 160,000 frames total)
Main datasets: Residual RL expert rollouts in Isaac Lab warehouse·office environments
Tasks & evaluation: Isaac Lab 4 envs×640 runs: Combined success rate X-Mobility→generalist Carter 50.3→85.3%, H1 22.0→84.4%, Spot 11.7→77.9%, G1 10.0→90.6%; specialists improve success rate 5–40× and weighted travel time 3× on average
Robot: Sim: Nova Carter wheeled, Unitree H1·G1 humanoids, Spot quadruped; real robots (v3): Carter·Unitree G1 (Jetson Orin, about 80% success over 20 random-obstacle runs)

Inference

Inference device & speed: AGX Jetson Orin TensorRT: P50 29.34 ms·P95 29.38 ms·GPU 422 MB
Inference memory: 422 MB (Jetson AGX Orin, TensorRT)
Weights bf16 (calc.): Not disclosed
Where it runs: Jetson AGX Orin (TensorRT)

Reinforcement learning & base

Reinforcement learning: Used (VLA RL post-training) · On top of the X-Mobility (imitation learning) policy, per-embodiment residual PPO RL fine-tunes the policy itself, then distills into a generalist; legged robots also combine a low-level RL locomotion policy
RL evidence: https://arxiv.org/html/2502.16372
Base control: Wheeled·Quadruped·Humanoid legged (Direct base velocity/pose output) · generalist policy outputs velocity commands (v, ω) → Carter drives directly; for H1·G1·Spot, Isaac Lab RL locomotion policy controls joints (sim)

Notes

Key features: Trains per-embodiment residual RL experts on top of an imitation-learning-based policy, then distills them into a single embodiment-conditioned policy to unify wheeled·quadruped·humanoid navigation. Raised success about 5× over X-Mobility on sim benchmarks (e.g., H1 22.0%→84.4%) and runs in 29 ms on Jetson AGX Orin.
Notes: The initial version mentioned real-world deployment only in the Figure 1 caption and left Sim2Real as future work, but v3 (2025-10) added real-robot deployment on Carter and Unitree G1 (Jetson Orin, about 80% success over 20 trials). HF (nvidia/COMPASS, gated) provides Carter ONNX and JIT policies and GR00T post-training materials; the multi-embodiment generalist checkpoint was not found in the file list. Provides a workflow for adding navigation to the GR00T VLA with distillation data (official docs). Parameters and training time not disclosed. NVIDIA Open Model License. Based on: X-MOBILITY: End-To-End Generalizable Navigation via World Modeling (arXiv 2410.17491) — X-Mobility is used as a frozen base policy (stated in the paper and HF card)
Verification: Corrections applied