ACE Robotics Publishes Kairos World Model, Reporting Four Embodied AI Benchmark Wins

ACE Robotics published the Kairos physical AI world model paper on arXiv, where it was picked up by Hugging Face Daily Papers. The 4B-parameter model uses hybrid linear temporal attention in a unified architecture. The team reported first place on RoboTwin 2.0 (96.1%), LIBERO-Plus (89.0), WorldModelBench Robot, and DreamGen, ahead of larger proprietary models. Weights are available under a permissive license on Hugging Face, GitHub, and ModelScope.

Visit original source