Researchers led by Yuxiang Yang have built PhysVAM, an action model for handling objects that keep moving while the robot looks and acts, according to a paper on arXiv. Instead of predicting future images, it predicts how upcoming interactions will help or hurt progress on the task.
The model learns this from successful and failed attempts over several time horizons and feeds it into the action policy. In simulation and real-world tests, it beat the strongest baselines by 27.0 and 27.5 percentage points in absolute success rate.



