Aug 27, 2026
arXiv preprint: Should a Compact Robot Controller Be Organized as Observation-to-Action Mapping at All? (PredVLA)
Is compressing a large VLA really the right way to build a small robot policy? PredVLA examines this question not by shrinking observation-to-action mapping, but from a different computational principle: predicting the sensorimotor world and inferring the latent state through prediction error. With no robot-data pretraining and about 0.68M trainable parameters, it reaches 86.94% mean success across the three short-horizon LIBERO suites, far above BC-Transformer (19.73%) and BC-LSTM (10.26%) given the same parameter budget and the same frozen front end.
Read more→

