Innovative Integrated Architecture: Kairos Redefines the Core Logic of World Models
Addressing common issues in existing world models, such as a lack of physical law adherence and insufficient causal reasoning, ACE ROBOTICS launched its self-developed Kairos World Model in December 2025. The core of the model is its pioneering “Multimodal Understanding-Generation-Prediction” integrated unified architecture, which completely abandons the “add-on” approach of post-training on top of video generation models.
Through a self-developed hybrid linear attention operator and a global state-sharing mechanism, this architecture deeply integrates the three core capabilities of environment understanding, future generation, and state prediction from the ground up. On the data front, Kairos was pre-trained on over 100,000 hours of human-centric real-world data and millions of hours of real-world internet videos. The training paradigm combines explicit imitation learning with latent space reinforcement learning, ensuring the model establishes a closed loop from data perception to deep understanding, significantly enhancing its scene recognition and temporal causal inference abilities.
Based on this architecture, the team developed the Kairos-4B model, the world’s first embodied world model capable of on-device deployment and direct robot actuation. It creates an end-to-end link from world understanding to state prediction, eliminating intermediate translation delays and laying the foundation for robots to perform high-real-time, high-precision “think-and-act” tasks.
Complex Manipulation and Scene Generalization: Topping RoboTwin 2.0 and LIBERO-Plus
The Kairos model demonstrated its leadership in complex task execution and adaptation to unknown environments in two key benchmarks.
On RoboTwin 2.0, a challenging dual-arm manipulation benchmark introduced by institutions like Shanghai Jiao Tong University and the University of Hong Kong, Kairos set a new record with a 96.1% average success rate, claiming the top spot. This result not only surpasses leading World Action Models (WAMs) like MotuBrain (96.0%) but also significantly outperforms Vision-Language-Action (VLA) models such as G0.5 (93.2%) and starVLA (88.3%), validating its exceptional capabilities in complex dual-arm coordination and fine-grained manipulation planning.
On LIBERO-Plus, the world’s most rigorous scene-level generalization benchmark, Kairos ranked first with a total score of 89.0. This benchmark, jointly proposed by Fudan University, the National University of Singapore, and others, tests model robustness by simulating seven real-world variables, including lighting, background, and noise. Kairos not only surpassed fellow world model Being-H0.7 (84.8) but also outperformed all leading VLA models like ACoT-VLA (88.0). Its near-perfect scores in dimensions such as lighting (97.7), background (95.8), and noise (96.8) demonstrate that robots equipped with this model can be directly deployed in diverse real-world settings like homes and factories with minimal adaptation costs.
Physics Modeling and Parameter Efficiency: Leading the WorldModelBench for Robotics
In the WorldModelBench for Robotics, an industry gold-standard test jointly launched by UC Berkeley, NVIDIA, and others, Kairos-4B topped the global leaderboard with a total score of 9.30 using only 4B parameters. This performance comprehensively surpasses several larger models, including the 28B-parameter Lingbot and the 16B-parameter Cosmos3, setting a new record for parameter efficiency in the world model domain.
This benchmark primarily evaluates a model’s instruction following and future frame generation capabilities, which are key metrics for fundamental embodied intelligence. Kairos scored 2.36 on instruction following, tying for first place with the 16B-parameter Cosmos3, representing a 4x improvement in parameter efficiency. In physics compliance, it achieved a score of 4.96, with perfect scores for its understanding of core physical laws like Newtonian mechanics and gravity. This result signifies that Kairos has reached a state-of-the-art level of precision in modeling the physical world.
Synthetic Data Engine: DreamGen Benchmark Validates Commercialization Value
Kairos achieved another breakthrough on the DreamGen Bench, a benchmark jointly launched by NVIDIA, the University of Washington, and others to evaluate the real-world generalization capabilities of world models. It secured the global top rank in both Average Physics Adherence (AVG_PA 0.538) and overall Average Score (AVG_Score 0.618).
Scores on this benchmark show a strong positive correlation with the actual performance of downstream robot policy training. Kairos significantly outperforms competitors in the physics adherence dimension for core generalization scenarios, such as executing new behaviors (PA 0.489) and adapting to new environments (PA 0.581). This means the synthetic data generated by Kairos not only adheres closely to physical laws but also transfers effectively to unseen objects, behaviors, and environments. This will drastically reduce the training costs and cycles for robots in new scenarios, providing a core data engine for the large-scale commercialization of embodied intelligence.
