Researchers led by Xu Pan have built CauSyncVLA, a controller that lets a vision-language-action policy keep acting while its visual processing is still running, according to a paper on arXiv. Running the two at once saves waiting time, but it leaves the robot's picture of the scene out of date.

A causal latent world model predicts the robot's current visual memory from delayed images and the states and actions already carried out. The actions still to come are then aligned with this memory without extra training. On LIBERO-10 it reached 93.8% success at 5.04 times the inference throughput.

Source: arXiv Robotics (cs.RO)

Background

Xu Pan

Robotics researcher

Robotics researcher.