At the 29th Information-Based Induction Sciences Workshop (IBIS2026), held October 26–29, 2026 at Sonic City in Saitama, I presented “PredVLA: Predictive Sensorimotor Modeling for Sub-Million-Parameter Robot Manipulation” as a poster (Poster Session 2, October 28). This is joint work with Shunichi Kasahara at Sony CSL.
Rather than compressing a small robot controller as an observation-to-action mapping, PredVLA organizes it around predicting sensorimotor dynamics and inferring its latent state from prediction error. With no robot-data pretraining and only about 0.68M parameters, it substantially outperformed baselines with the same parameter budget on the LIBERO benchmark.