机器人不能停下来等模型：星尘发布 SmoothRL，让在线强化学习跟上大模型的异步推理

TL;DR - Astribot introduced SmoothRL, an online reinforcement-learning framework that trains robot policies under the same asynchronous inference conditions used in deployment. By updating only actions actually executed, it improves real-world manipulation accuracy and continuity without forcing robots to pause for model inference.

• SmoothRL divides action chunks into committed, executed, and discarded regions, allowing gradients to pass only through the executed region.
• Training rollouts run inference and physical execution concurrently, aligning reinforcement learning with real deployment timing.
• On Astribot’s S1 robot, success rates improved from 39% to 94% for throwing, 8% to 83% for pen capping, and 30% to 90% for package opening.
• Smoothness constraints reduced end-effector acceleration RMS by 52% and jerk by 47% in a real throwing rollout.