What Should Robot Policies Learn to Predict? Rethinking Supervision Targets in Vision-Language-Action Learning
The study tested whether allocating training samples to auxiliary prediction tasks, like latent-action (LA) and future-vision (FV), could improve robot policy learning compared to focusing solely on direct action prediction. The results show that suitable LA and FV targets can outperform action-only training under the same budget, suggesting that explicit design choices for predictive targets are better than defaulting to action supervision.