Robotics preprint reports lower video bandwidth and higher task success

Researchers report a video compression method that uses a robot's known motion to reduce the camera bandwidth needed by remotely hosted control models. In a preprint, they report 23.9% lower bit rate than a capacity-matched codec conditioned on video history, and 40-56% lower bit rate than x264, x265 and SVT-AV1 on LIBERO. Both ends use joint states, the robot model and camera calibration to predict image changes without transmitting an additional conditioning signal. At 6 kbps, OpenVLA-OFT task success rose from 22.8% to 43.1%; pi_0.5 success rose from 55.5% to 77.4%. These task results come from benchmark evaluations; the authors separately report transfer to real robot data.