DreamTrue robot world model cuts interaction defect rate from 48.12% to 6.25%, tops AgiBot challenge
Junyan Li · hf · 2026-10-09
- DreamTrue is a multi-view, cross-embodiment robot world model for action-faithful, physically plausible video prediction; it ranked first in the world model track of the AgiBot World Challenge 2026.
- Fixes two training obstacles: offline geometric calibration aligns image-space action-trajectory conditions to improve action following; counterfactual post-training regenerates futures under modified actions/contacts to widen interaction coverage.
- A human-annotated defect dataset trains an embodied video reward model that guides RL post-training.
- Results: state-of-the-art action following on AgiBot, human-assessed interaction defect rate cut from 48.12% to 6.25%.
More from Embodied
- Scoble rides Tesla Cybercab: exceeds expectations, plans to buy one to rent out — Scobleizer · 2026-10-09
- Egocentric data collection is all the rage now, says researcher citing an ahead-of-its-time project — gragtah · 2026-10-09
- NVIDIA's USDCraft uses LLM-written programs to build simulation-ready articulated 3D assets — nvidia · 2026-10-09
- SuperNav: ZJU's Agent Harness Lets MLLMs Navigate Any Scene Without Fine-tuning — zju3dv · 2026-10-09
- SimpleICL Defines Robot In-Context Learning with a Low-Cost Open Recipe — Minxing Li · 2026-10-09
- UBTECH inks FAW-Volkswagen deal for factory humanoids, targets 10,000 robots a year — CyberRobooo · 2026-10-09