DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training
DreamTrue cuts human-rated robot interaction defects from 48.12% to 6.25% on AgiBot.
The paper presents a robot world model aimed at making predicted videos follow actions more faithfully and obey contact physics more plausibly. It uses image-space action trajectories with offline geometric calibration to handle imperfect robot dataset alignment. The authors also add counterfactual post-training, generating futures from modified action trajectories and scoring them with a human-trained embodied video reward model. They report state-of-the-art action following on AgiBot and first place in the AgiBot World Challenge 2026 world model track. HF Daily Papers' note
The paper presents a robot world model aimed at making predicted videos follow actions more faithfully and obey contact physics more plausibly. It uses image-space action trajectories with offline geometric calibration to handle imperfect robot dataset alignment. The authors also add counterfactual post-training, generating futures from modified action trajectories and scoring them with a human-trained embodied video reward model. They report state-of-the-art action following on AgiBot and first place in the AgiBot World Challenge 2026 world model track. HF Daily Papers' note
score 4