ACE ROBOTICS has announced that its open-source Kairos world model has achieved leading results across four global embodied-intelligence benchmarks: RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot and DreamGen. As of June 12, 2026, Kairos ranked first among evaluated world models and vision-language-action (VLA) systems on these benchmarks' public leaderboards, leading across core capabilities including complex robotic manipulation, scene-level generalization, physical-world modeling and zero-shot transfer. The project is openly available on GitHub, Hugging Face and ModelScope, providing researchers and developers with public access to the model, benchmark results and technical materials.
Embodied intelligence faces a fundamental challenge: generalization. Robots must operate reliably in unseen environments, adapting to new lighting, layouts, objects, embodiments and noisy real-world conditions. While VLA models have been a prevailing approach by directly mapping perception and language inputs to robot actions, ACE ROBOTICS believes world models offer a more scalable path by explicitly learning the underlying dynamics of the physical world. Kairos is designed to validate that approach.
One of Kairos' most significant results comes from LIBERO-Plus, a scene-level generalization benchmark proposed by the Shanghai Innovation Institute with Fudan University, Tongji University and the National University of Singapore. It evaluates robustness under seven real-world variables: camera angle, robot embodiment, language instruction, lighting, background, sensor noise and spatial layout. Kairos achieved an overall score of 89.0, ranking first among all evaluated world models and VLA systems, surpassing leading VLA models including ACoT-VLA (88.0), Pi 0.5 (85.7) and ProGAL-VLA (85.5), as well as the Being-H0.7 world model (84.8). It showed strong environmental robustness, with near-ceiling performance on lighting (97.7), noise (96.8) and background (95.8). According to ACE ROBOTICS, this marks the first time a world-model approach has outperformed leading VLA systems on LIBERO-Plus for scene-level generalization, pointing to a path where robots adapt to homes, factories, retail spaces and other environments with far less environment-specific retraining.
On WorldModelBench Robot, a physical-modeling benchmark proposed by researchers from UC Berkeley, UC San Diego, NVIDIA and MIT, Kairos-4B achieved an overall score of 9.30, ranking first. With only 4 billion parameters, it outperformed larger systems including 28-billion-parameter Lingbot, 16-billion-parameter Cosmos 3, 14-billion-parameter Abot-PhysWorld and 5-billion-parameter Wan 2.2, setting a new record for parameter efficiency. Kairos matched the top instruction-following score (2.36) of the 16-billion-parameter Cosmos 3 with about one quarter of the parameters, a fourfold efficiency gain. It scored 4.96 on physics adherence, with perfect marks on Newtonian mechanics and gravity, and a perfect score on temporal quality.
ACE ROBOTICS attributes Kairos' performance to its native unified "multi-modal understanding-generation-prediction" architecture, which integrates world understanding, generation and prediction within a single backbone that shares one global world state. This reduces information loss and coordination latency for more consistent physical modeling, stronger long-horizon prediction and more reliable action planning. ACE ROBOTICS first introduced this architecture in December 2025, and the broader industry is now converging on a similar path: NVIDIA's Cosmos 3.0, introduced in 2026, adopts a comparable single-system design. Built on this foundation, Kairos-4B is the first embodied world model able to drive a physical robot directly on-device, closing the perception-to-action loop without intermediate translation latency.
Kairos also ranked first on DreamGen Bench, a benchmark led by NVIDIA with the University of Washington, UC Berkeley and UCLA, measuring how well synthetic data generated by world models transfers to unseen objects, behaviors and environments. Kairos ranked first on both average physics adherence (AVG_PA 0.538) and overall average score (AVG_Score 0.618), and led globally on new-behavior execution and new-environment adaptation. On RoboTwin 2.0, a dual-arm manipulation benchmark proposed by Shanghai Jiao Tong University and the University of Hong Kong with Shanghai AI Laboratory, Kairos scored 96.1% — a state-of-the-art result. Across the benchmark's 50 complex two-arm tasks it scored 96.9% in clean scenarios and 95.2% in randomized scenarios, ahead of VLA models such as G0.5 (93.2) and starVLA (88.3) and world models including AIM (93.1), Fast-WAM (91.8) and MotuBrain (96.0).
Together, these results validate Kairos' technical direction across the core dimensions of embodied intelligence, supporting ACE ROBOTICS' aim to move robots beyond task imitation toward physical-world understanding, long-horizon reasoning and real-world execution. The results come as ACE ROBOTICS accelerates commercialization, having raised several hundred million U.S. dollars across financing rounds in the first half of 2026, including a recent Angel+ round backed by investors such as Dachen Caizhi, Shenzhen Capital Group and the Shanghai Sci-Tech Innovation Fund, with existing shareholder SenseTime's Guoxiang Capital increasing its stake. The proceeds will support continued world-model research and integrated hardware-software solutions for sectors including smart retail, security and 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."
