training-signal allocation for vision-language-action robot policiesExpert
What Should Robot Policies Learn to Predict? Rethinking Supervision Targets in Vision-Language-Action Learning
Robot AI models are usually trained almost entirely on predicting the next motor command. This preprint from Renmin University, Peking University, Tsinghua and the model maker ModelBest shows that spending half or more of the same training budget on predicting what the camera will see next produced better robots in simulation and on four real tasks.