Kairos (ACE Robotics)

A 4-billion-parameter embodied world model from ACE Robotics that does understanding, video generation, and action prediction all at once.

Kairos is the world model of Shanghai-based embodied-AI company ACE Robotics (Daxiao Robotics); its technical report lists Xiaogang Wang and Dacheng Tao among the authors, and the company is reportedly led by Xiaogang Wang, a co-founder of SenseTime. Kairos 3.0 was released in December 2025 with 4-billion-parameter pretrained weights open-sourced; a technical report followed in June 2026, and version 3.1 plus world-action-model inference code was open-sourced in July. Its central claim is that a world model doesn't need to render every pixel photorealistically — what matters is retaining the information useful for control, such as object state, contact, task progress, and the consequences of an action. Concretely, it orders training data in stages — ordinary video first, then human behavior, then robot interaction — uses a unified architecture for understanding, generation, and prediction all at once, and uses hybrid linear temporal attention to cut the cost of long-horizon inference.

ExampleThe open-sourced kairos-4B-robot-RoboTwin2.0 weights jointly predict future frames and actions across more than 50 bimanual tasks on RoboTwin 2.0, achieving what the company says is the best result yet on that benchmark.

Also called

A Native World Model Stack for Physical AI, A Regret-Aware Native World-Action Model Stack for Physical AI

Related

World Model · World Action Model · Embodied Foundation Model · Video Generation Model · ACE Robotics · RoboTwin

Sources

Kairos: A Regret-Aware Native World-Action Model Stack for Physical AI (arXiv 2606.16533，v1 题为 A Native World Model Stack for Physical AI) (Chinese)
kairos-agi/kairos (GitHub)
大晓机器人官网 (Chinese)

As of

2026-07