FACT turns robot failures into training data for smarter world-action models

World-action models (WAMs) typically co-train policies with future prediction to give robots physical priors for action generation. But they learn mostly from successful demonstrations, so they have no incentive to predict the consequences of bad actions. This creates a success bias that blinds robots to failure modes. FACT, a new causal World-Action Model from UC San Diego and KAIST, solves this by conditioning future video and task-progress predictions on the executed action. That action-conditioned interface lets failure rollouts supervise action consequences, turning bad actions into valid future targets instead of discarding them. The progress predictor becomes aware of both successful and failed outcomes, and can also score sampled action candidates at inference for extra robustness.

In extensive tests on simulation and real-world bimanual manipulation tasks, FACT outperformed several existing baselines, with accuracy improving as more failure data was added to training. It also reduced success-biased future hallucination when the model encountered bad actions. The approach is particularly useful for safety-critical robotic applications where predicting what goes wrong matters as much as knowing what works. By leveraging failure data that most pipelines throw away, FACT points toward a more data-efficient and reliable training paradigm for embodied AI.