DreamTrue Unveils Robot World Model with Counterfactual Fine-Tuning

Researchers published a paper on arXiv, DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training, in which they presented a multi-view robot world model for predicting action-consistent video. The model was trained on AgiBot data and took first place in the world model track of the AgiBot World Challenge 2026 competition. The authors released the code as open source.

What it means

DreamTrue is a world model for robotics, not a video generator for user tasks. In our video generators section there are no tools that solve the task of predicting future frames from robot action trajectories. The closest tools in form — Maestro, LTX-Video and CogVideoX — work with text or visual prompts, not with action trajectories and contact configurations. The authors of DreamTrue specify neither the model size, nor VRAM requirements, nor the license. All that is known is that the code is open and that training and evaluation were carried out on AgiBot data. For comparison: LTX-Video claims a requirement of 1 GB of VRAM, Maestro — 12 GB, and CogVideoX targets NVIDIA H100 and above. There is no practical overlap with our reference yet: DreamTrue solves the task of physically plausible prediction of robot interactions with objects, not generation of video from text or an image. If the authors publish the weights and hardware requirements, it will be possible to assess how suitable the model is for running on your own hardware.