AI
CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 03:33:27 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 03:18:25 UTC.
Briefings
CrossBFM addresses the high GPU-hour cost and embodiment-specific latents in Behavior Foundation Models by proposing a unified encoder architecture with no robot-specific parameters. This approach distills a shared latent behavior space across three humanoids using only 10 GPU-hours total, enabling all prompting modes to transfer effectively. Experiments show that regressing the encoder on a quarter of the motion corpus costs only 5% of tracking performance, and training on a subset of robots recovers up to 89% of tracking performance on unseen morphologically similar robots.
Key points
- CrossBFM uses a unified encoder architecture with no robot-specific parameters to distill a shared latent behavior space across multiple humanoids, reducing training costs from hundreds of GPU-hours to under 10 GPU-hours.
- Latent-conditioned trackers successfully transfer motion tracking, smooth goal reaching, and reward optimization across three humanoids, with only 0.025 rad loss in motion tracking compared to joint-conditioned counterparts.
Context and limitations
The pipeline was verified on real robots across all three prompting modes and with flow-based generated latents, but the specific hardware configurations and exact real-world deployment conditions for the three distilled humanoids are not detailed in the evidence.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 03:33:27 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeicsdb6wy4v62z6eyiqp7sa65bjxxuv6ocfizs53x245a5qnas7yp4