Researchers propose a video codec that uses a robot’s joint states, model and camera calibration to predict how its own movement changes the camera view. The shared motion signal adds no video bits, and helps the codec send less footage over a limited-bandwidth connection.
In LIBERO tests, the researchers report that OpenVLA-OFT task success rose from 22.8% to 43.1% at 6 kbps, compared with a history-conditioned codec. The work is a preprint; its authors say code and supplementary material will be released after formal publication.
Why it matters
For teams running robot policies on remote servers, the reported gains suggest motion-aware video compression could improve task success when camera uplink bandwidth is tight; real-world performance remains to be established.
