DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training

DreamTrue. Action-Faithful Robot World Model with Counterfactual Post-Training. By Junyan Li and eight others, from the Institute of Automation, Chinese Academy of Sciences and Amap, Alibaba Group. Published on the eighth of October.

A robot world model predicts video of what will happen when the robot acts. For that prediction to be useful it has to follow the action faithfully and stay physically plausible. Training on existing robot datasets runs into two obstacles. Imprecise calibration can impair action following. And because these datasets cover few unsuccessful interactions, predictions can be biased toward success. A model that has mostly watched interactions succeed will lean toward predicting success, whatever the action actually says. Neither problem is about image quality. Both are about whether the predicted future is the one the given action would really produce.

DreamTrue is a multi-view, cross-embodiment robot world model that answers each obstacle in turn, and adds a learned judge of physical plausibility.

For action following across embodiments, action trajectories are rendered into image-space conditions, and an offline geometric calibration aligns those conditions with the target videos. The action is shown to the model in the same space as the pictures it predicts, and the calibration makes sure the two line up.

For coverage, counterfactual post-training modifies recorded action trajectories and generates future videos under a wider range of actions and contact configurations. Those predictions have no paired ground-truth future to compare against. To give feedback anyway, the authors build a human-annotated video dataset of robot, object and interaction defects, and train an embodied video reward model on it. Its scores then guide reinforcement-learning post-training toward more physically plausible outcomes.

On AgiBot, DreamTrue reaches state-of-the-art action following. The human-assessed rate of interaction defects falls from forty-eight point one two percent to six point two five percent. The model also ranked first in the world model track of the AgiBot World Challenge two thousand and twenty-six. The defect figure is the one that speaks to the second obstacle. Before, nearly half of the assessed predictions showed a defect in the interaction. After, only a small fraction do.

The training data only ever showed what robots did, and mostly what went well. Counterfactual post-training lets the model see what other actions would have produced, and the reward model supplies the judgment that real footage of those futures would otherwise have had to provide.