New robot model hits 88% where pi-0.5 gets 1%
A Washington group says grounding objects in 3D before learning actions takes a swapped-target benchmark from 1% to 88%. Two other results landed too, both about making big models usable on real robots.
Shown this week
Grounded Action Model beats pi-0.5 when the target object changes
Jiafei Duan's group introduced GAM, which builds a robot policy on top of a frozen pretrained 3D grounding model and trains only the action head. On LIBERO-PRO it reports 61% average success across 16 settings versus 53% for pi-0.5, and when the target object changes it reports 88% versus 1%.
Ground first. Then learn to act.
Berkeley and Siemens figure out how to do RL when the robot lags
A Berkeley and Siemens team released ARLI, a way to reinforcement-learn a VLA (a model turning camera images and text into robot actions) even when inference delay means actions come from stale observations. Sergey Levine says the trick is a small, fast RL policy that sees more recent images and steers a large robot foundation model.
we figured out how to use a small RL policy with a large robot foundation model
Astra hits 79% on RoboMME at 3.63 calls per episode
Bingao Chen reports 79.13% success on RoboMME using GPT-6 Astra with just 3.63 Astra calls per episode, evaluated on 800 official test episodes across 16 tasks. The setup is three tiers: Astra for reasoning, a small visual monitor, and a VLA for the actions.
let Astra decide what to do, and let the small model decide when to ask it again