DreamTrue
A video world model for robot learning that improves action fidelity and physical plausibility through multi-view, cross-embodiment video prediction, counterfactual post-training, and an embodied video reward model. It aims to produce future video consistent with the commanded actions and the physical constraints of the environment.
Timeline (last 90 days)
DreamTrue is a method that realizes a robot world model through a multi-view, cross-embodiment video prediction framework, focusing on improving action faithfulness and physical plausibility.
DreamTrue renders action trajectories into image space to form visual conditions, then aligns these conditions with the target video through offline geometric calibration, so that the cross-embodiment model better understands action intent.
DreamTrue introduces counterfactual post-training, which generates future videos under a broader range of action and contact configurations by modifying recorded action trajectories, in order to broaden interaction coverage.
DreamTrue builds an embodied video reward model to score the physical plausibility of predicted videos, and the scores are used for reinforcement learning post-training.
On the AgiBot platform, DreamTrue achieves state-of-the-art action-following performance.
DreamTrue reduced the manually evaluated interaction defect rate from 48.12% to 6.25%, an improvement of more than 80%.
DreamTrue ranked first in the world model track of the AgiBot World Challenge 2026.
All Facts (7)
DreamTrue is a method that realizes a robot world model through a multi-view, cross-embodiment video prediction framework, focusing on improving action faithfulness and physical plausibility.
50%UnverifiedDreamTrue renders action trajectories into image space to form visual conditions, then aligns these conditions with the target video through offline geometric calibration, so that the cross-embodiment model better understands action intent.
50%UnverifiedDreamTrue introduces counterfactual post-training, which generates future videos under a broader range of action and contact configurations by modifying recorded action trajectories, in order to broaden interaction coverage.
50%UnverifiedDreamTrue builds an embodied video reward model to score the physical plausibility of predicted videos, and the scores are used for reinforcement learning post-training.
50%UnverifiedOn the AgiBot platform, DreamTrue achieves state-of-the-art action-following performance.
50%UnverifiedDreamTrue reduced the manually evaluated interaction defect rate from 48.12% to 6.25%, an improvement of more than 80%.
50%UnverifiedDreamTrue ranked first in the world model track of the AgiBot World Challenge 2026.
50%