CrossBFM distills a shared latent behavior space across humanoid robots in under one GPU-hour
Jan_R_Peters · x · 2026-10-01
- New work CrossBFM (Tan-Dzung Do, Nico Bohlinger, Jan Peters, et al.) treats the behavior latent space itself as the transferable asset across humanoid embodiments.
- Existing Behavior Foundation Models use Forward–Backward representations to give one robot a promptable policy (a latent vector specifying motion imitation, goal reaching, or reward maximization), but cost hundreds of GPU-hours per robot and yield incompatible latent spaces across robots.
- CrossBFM exploits retargeting-induced frame-level correspondence, turning latent transfer into supervised regression with no simulator or training on the target side. A unified encoder with no robot-specific parameters serves all embodiments in under 1 GPU-hour, and latent-conditioned trackers produce whole-body control via standard PPO in 10 more GPU-hours.
- Validated on three humanoids: motion tracking within 0.025 rad of joint-conditioned policies, smooth fall-free goal reaching, and transferable reward optimization. Code coming soon.
Related event: CrossBFM distills shared behavior latent space across humanoid robots(2 posts)→
More from Embodied
- 5 robotics accounts that are '6 months ahead': Jim Fan, Physical Intelligence founders lead the list — RemiCadene · 2026-10-01
- DyRAD: radar novel view synthesis for dynamic driving scenes, recovering 90.7% vs 26.9% baseline — Technion · 2026-10-01
- VLA reinforcement learning updates are low-rank and concentrate in overlooked Timestep Modules — HOLILAB · 2026-10-01
- Huawei Mate 90 debuts Kirin 9050 Pro, first chip with LogicFolding architecture — ingliguori · 2026-10-01
- Paper deadline in 24h: researcher built a LeRobot SO-101 rig in an Airbnb and trained policies in hours — RemiCadene · 2026-10-01
- Raise Robotics' fastening robot puts up high-rise walls so workers never leave the ground — lukas_m_ziegler · 2026-10-01