CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments
CrossBFM distills one latent behavior space that transfers across humanoid robots in under an hour of GPU training.
CrossBFM distills a shared latent behavior space across humanoid embodiments so one prompt vector can request imitation, a target pose, or reward maximization. A unified encoder with no robot-specific parameters trains on all embodiments in under a GPU-hour, and latent-conditioned PPO trackers then learn whole-body control in about 10 more GPU-hours. On three humanoids, tracking loses only 0.025 rad versus joint conditioning, goal reaching has no falls, and all 41 reward prompts optimize. Training on a robot subset and testing an unseen similar robot recovers up to 89% of seen tracking performance, and the pipeline is shown on real robots.
- Shared encoder has no robot-specific parameters and trains in under one GPU-hour.
- Latent trackers then train with PPO in about 10 additional GPU-hours.
- Motion tracking loses only 0.025 rad versus a joint-conditioned policy.
- An unseen similar robot recovers up to 89% of seen-robot tracking performance.
- Verified on real robots for tracking, goal reaching, and reward prompts.
- AI models
- CrossBFM
From the original article
1 min read at huggingface.co
Behavior Foundation Models (BFMs) give humanoids a promptable policy over a latent behavior space, enabling one single vector to represent a motion to imitate, a pose to reach, or a reward to maximize. Forward-Backward representations successfully produce such spaces, but at the cost of hundreds of GPU-hours for a single robot.
By Tan-Dzung Do, Tuan Dat Phuong, Nico Bohlinger · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.