Astribot's SmoothRL: online RL learns only from actions the robot actually executes

jiqizhixin · x · 2026-09-12

Asynchronous inference is now the norm in real robot deployment: while the robot executes action A, the model already computes action B. But models generate whole action chunks at once, and only a fraction gets executed — planned and executed actions diverge. What should online RL learn from?

Astribot presents SmoothRL (Online Reinforcement Learning During Asynchronous Execution), which splits each action chunk into three regions: Committed (submitted, will execute), Execution (frames truly executed this round), and Discarded (overwritten by the next chunk before execution). Only the Execution Region passes value gradients back, so the policy learns from what the robot actually did — not planned-but-never-executed actions.

Original post →

More from Embodied

Embodied channel →