The Gaussian is enough: Toyota study finds non-Gaussian priors don't help fine-tuning LBMs
_krishna_murthy · x · 2026-09-25
A negative-result study from Toyota Research Institute, Woven by Toyota and Cornell ("The Gaussian Is Enough") shows that closer-to-target non-Gaussian action priors—which help when training imitation policies from scratch—offer no benefit when fine-tuning pretrained Large Behavior Models. Across 100K+ simulation rollouts on three LBM architectures (LBM 1.0, π0.5, GR00T N1.5), 40+ tasks, and 1250 hardware rollouts on five bimanual tasks, the standard Gaussian prior performed statistically indistinguishably or better, with a possible exception at very low fine-tuning data fractions. Diagnostic analyses indicate the observation encoder, not the action prior, governs fine-tuning performance.
More from Embodied
- NYU's Steertuning Lets a Robot Learn a Chore From 5 Minutes of Human Demo — CyberRobooo · 2026-11-11
- Boston Dynamics Atlas demoed with Blender 5.3 — TomLikesRobots · 2026-11-11
- Lamport Robotics posts robot 'wake-up test' demo video — R4nsack · 2026-11-11
- Embodied AI Faces Two-Order-of-Magnitude Data Gap as Industry Heads to 10M Hours — GaryMarcus · 2026-11-11
- NVIDIA hosts CoRL 2026 meetup on scaling robot training with Isaac Lab — m_wulfmeier · 2026-11-11
- Lighthouse extends Obsidian welding foundation model to run on any robotic platform — bennash · 2026-11-11