· Industry· 日本語で読む
NVIDIA Streamlines Cross-Embodiment Locomotion Policy Training With COMPASS Workflow
NVIDIA on Aug. 26, 2026, unveiled COMPASS, an agent-driven workflow for training cross-embodiment locomotion policies from a single-embodiment demonstration. Using Boston Dynamics’ Spot as the reference robot, it trains a specialized policy with residual reinforcement learning on top of the pretrained X-Mobility policy. The workflow supports bundled warehouse scenes and SAGE-10K-generated scenes, and optionally can reconstruct real environments with NVIDIA Omniverse NuRec.
What happened
NVIDIA on Aug. 26, 2026, unveiled COMPASS, an agent-driven workflow. COMPASS trains cross-embodiment locomotion policies from a single-embodiment demonstration. Using Boston Dynamics’ Spot as the reference robot, it reuses the pretrained NVIDIA X-Mobility policy and trains a specialized policy with residual reinforcement learning.
Why it matters
The workflow is aimed at developers who want to create locomotion policies for other robots from a single demonstration. NVIDIA outlines a process for reusing pretrained X-Mobility and adapting it to new robots and environments with residual reinforcement learning. It details conditions for progressing through validation, training and evaluation using Spot as the reference robot, with bundled warehouse scenes, SAGE-10K scenes or environments created with Omniverse NuRec.
Sources
Translated from the Japanese original: /news/nvidia-robotto-no-nabigeeshon-gakushuu-wo-jidouka-suru-eejie-1900a6. The Japanese version is authoritative if they differ.