Gaussian head on flow policies: 0.02% of data, similar success, up to 2x faster inference
francoisfleuret · x · 2026-10-30
Researchers show flow policies can be approximated by a Gaussian head fit on just 0.02% of data: similar task success rates, up to 2x faster inference, and closed-form likelihoods. François Fleuret explains that real-world action planning needs a large continuous action space and complex posterior — but as with diffusion models, at a small time scale a plain Gaussian suffices, with behavioral diversity preserved via diffusion in physical time. Evaluated on Xiaomi-Robotics-1 and GR00T.
More from Embodied
- Subvocal's YC-backed glasses decode silent speech at 200+ wpm, built in 4 months — garrytan · 2026-10-31
- Nomagic warehouse robots go from bimanual to trimanual, pause to solve tricky picks — m_wulfmeier · 2026-10-31
- Ben Goertzel: an AGI breakthrough could compress warehouse robotization from years to months — bengoertzel · 2026-10-31
- Waymo adds a witch hat Easter egg to its rider app for Halloween — jonathanmendez · 2026-10-31
- Figure paid people $15M to film chores, lifting robot zero-shot success in new homes from 9% to 56% — Scobleizer · 2026-10-31
- UMI-Live open-sources bimanual robot data collection; π0.5 hits 10/10 on real task — DJiafei · 2026-10-31