ACE ROBOTICS has announced that its open-source Kairos world model has achieved leading results across four major embodied-intelligence benchmarks, outperforming both vision-language-action (VLA) systems and rival world models in complex robotic manipulation, scene-level generalisation, physical-world modelling and synthetic data transfer.
As of 12 June 2026, Kairos ranked first on the public leaderboards of RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot and DreamGen – a sweep that its backers say validates world models as a more scalable path to general-purpose robotics.
On LIBERO-Plus, a scene-level generalisation benchmark that tests robustness against seven real-world variables – including camera angle, lighting, background, sensor noise and spatial layout – Kairos achieved an overall score of 89.0, surpassing leading VLA models such as ACoT-VLA (88.0) and Pi 0.5 (85.7).
It posted near-ceiling scores on lighting (97.7), noise (96.8) and background (95.8). ACE ROBOTICS notes this is the first time a world-model approach has beaten VLA systems on this benchmark for scene-level generalisation, pointing to reduced need for environment-specific retraining in homes, factories and retail spaces.
The model also demonstrated striking parameter efficiency. On WorldModelBench Robot – a physical-modelling benchmark from researchers at UC Berkeley, UC San Diego, NVIDIA and MIT – the Kairos-4B variant scored 9.30 overall, ranking first with only 4 billion parameters.
It outperformed much larger systems, including Lingbot (28B parameters) and Cosmos 3 (16B), and matched the top instruction-following score of Cosmos 3 with roughly one-quarter of the parameters.
ACE ROBOTICS attributes the performance to a native unified “multi-modal understanding-generation-prediction” architecture, which integrates perception, world dynamics and action planning into a single backbone – a design the broader industry is now converging on, including NVIDIA’s recent Cosmos 3.0.
Kairos also led on DreamGen Bench, a synthetic-data benchmark led by NVIDIA and multiple universities, ranking first on average physics adherence (0.538) and overall average score (0.618). On RoboTwin 2.0, a dual-arm manipulation benchmark from Shanghai Jiao Tong University and the University of Hong Kong, Kairos scored 96.1%, ahead of both VLA models and other world models.
The results arrive as ACE ROBOTICS accelerates commercial deployment, having raised several hundred million US dollars in the first half of 2026. Proceeds will back world-model research and integrated hardware-software solutions for smart retail, security inspection, tourism and hospitality.
“Embodied intelligence is the next era of AI, and a world model is the key to unlocking it,” said Wang Xiaogang, Chairman of ACE ROBOTICS. “Our mission is to give every robot a capable brain.” The Kairos project, including model weights and technical materials, is openly available on GitHub, Hugging Face and ModelScope.
