Event

Delayed Randomization Broadens Latency Margin of Humanoid Motion Trackers

1 reports · 1 sources

Research by Shaoqin Li and Yi Sun shows that randomizing command delays during training can significantly improve the tolerance of learned humanoid motion tracking policies to latency. The study tested a 24-joint humanoid robot in an independent CPU MuJoCo environment across five motions based on monocular video. After randomizing delays to 0–50 ms during training, the average survival rate under a constant 40 ms delay rose from 15% (0%–45%) to 95% (75%–100%), and the latency margin expanded from 20–30 ms to 50–70 ms. The cost is slightly looser tracking accuracy: errors for four motions increased by 0.02–0.06 rad, and for the most dynamic motion, robustness to other perturbations decreased. The authors accordingly recommend setting the randomization range based on the robot's measured latency and listing latency margin as one of the evaluation metrics for sim-to-real transfer.