Event

Future Video Prediction Improves Robot Action Learning Success Rate

1 reports · 1 sources

Hanwen Cui and other authors conducted controlled ablation experiments on two world action models, FastWAM and OpenWAM: keeping the model architecture, initialization, robot training data, and training steps fixed, they changed only the future video prediction loss. The results show that under clean conditions on seen tasks, adding prediction supervision increased success rates by 24.38 and 72.91 percentage points, respectively. However, this gain is not stable under distribution shift: on unseen tasks, the improvement varies by model and task; under the environment randomization conditions evaluated by the authors, both models had low success rates, and prediction supervision also showed no consistent advantage. The authors also introduced future video prediction as an auxiliary training objective into π0.5, using a lightweight Transformer predictor only during training, and the success rate rose from 82.94% to 89.19%.