Turns out robot flow policies are Gaussians in disguise 🥸🔔
A tiny Gaussian head, nearly the same success, and up to 2× faster inference.
Really fun project!
A tiny Gaussian head, nearly the same success, and up to 2× faster inference.
Really fun project!
1/8) Flow policies are Gaussian 😮
We replace their last layer with a Gaussian head fit on 0.02% of data:
• Similar task success
• Up to 2x faster inference
• Closed-form likelihoods
And behavior can still be diverse via diffusion in physical time.
(Xiaomi-Robotics-1, GR00T)View quoted post on X
