ACE Robotics Unveils Kairos World Model with Record Performance on Four Global Benchmarks
SHANGHAI, July 22, 2026 (GLOBE NEWSWIRE) -- ACE Robotics announced that its self-developed Kairos World Model has ranked first across four leading global embodied AI benchmarks: RoboTwin 2.0, LIBERO-Plus, WorldModelBench Robot, and DreamGen. The model also achieved multiple new state-of-the-art (SOTA) results. Together, these benchmarks evaluate bimanual manipulation, long-horizon task planning, text-driven embodied generation, and physical interaction reasoning. The results validate Kairos World Model's leadership in scene understanding, action prediction, physical consistency, and cross-task generalization for real-world robotics.
On RoboTwin 2.0, an international benchmark for bimanual robotic manipulation, Kairos World Model delivered superior simulation quality and stronger sim-to-real transfer. Its high-quality training data significantly improved zero-shot performance on real dual-arm robot tasks.
On LIBERO-Plus, which measures long-horizon robotic planning, the model maintained consistent scene understanding throughout complex multi-step tasks. It also avoided common issues such as object deformation and broken task logic.
On WorldModelBench Robot, Kairos World Model ranked first in text-to-visual generation and dynamic scene prediction. It accurately translated natural language instructions into realistic robot behaviors while preserving physical consistency.
On DreamGen, developed by NVIDIA GEAR Lab, Kairos World Model led both Physical Alignment (PA) and Interaction Fidelity (IF). Its generated motions followed realistic physical dynamics, producing simulations that can more effectively support real-world robotic deployment.
Unlike the widely used industry approach of combining a general-purpose video generation model with continued training and fine-tuning on embodied scenarios, the Kairos World Model follows a fundamentally different technical path. In December 2025, Ace Robotics introduced what it described as the world's first natively integrated architecture for multimodal understanding, generation, and prediction. The architecture brings world perception, scene generation, and action decision-making into a single unified backbone network, creating a complete intelligence loop from understanding the world to predicting it and interacting with it. In March 2026, the architecture was successfully validated for edge deployment. Nvidia's subsequently released Cosmos 3.0 adopted an architecture based on the same underlying design, indicating that the technical system pioneered by Kairos has become an important direction for the development of world model industry standards.
In terms of performance, Kairos World Model 3.0 supports long-duration video generation of up to seven minutes, while producing high-fidelity static and dynamic scenes at inference speeds that exceed those of mainstream world models. Its lightweight version, Kairos 3.0-4B, runs 72 times faster than Cosmos 2.5 on an A800 GPU, setting a new global performance record for embodied world models. More importantly, Kairos 3.0-4B is the world's first world model capable of driving embodied robot control at the edge. It is also the first embodied world model in the industry to achieve a 1:1.5 ratio of video generation time to video duration on the THOR edge platform, enabling real-time generation. Robots powered by the model can carry out complex tasks requiring long-duration and long-horizon reasoning, including preparing breakfast, performing household chores, and watering plants. The model also addresses widely recognized technical challenges such as fluid interaction and deformable-object manipulation. Its "one brain, multiple embodiments" capability allows it to adapt seamlessly to robotic hands, humanoid robots, robotic arms, and other hardware platforms from different manufacturers.
At the application level, Ace Robotics, CUHK MMLab, and Shenzhen Loop Area Institute jointly released Kairos-HomeWorld, the world's first whole-home interactive 3D world model. For the first time, the model enables end-to-end generation of a complete residential 3D environment from a single text prompt, with globally consistent structures, physically plausible layouts, and interactive objects. The research team also open-sourced the world's largest whole-home 3D dataset and the first designed specifically for Chinese households. It includes 300,000 floor plans from real Chinese homes and 5,000 complete simulated environments with interactive furniture and objects. Kairos-HomeWorld is now used in Ace Robotics' routine embodied intelligence training, supporting full-process simulation training for complex, long-horizon household tasks such as cross-room navigation and multi-room object organization.
From generating video to controlling physical robots, and from single-scene generation to fully interactive whole-home environments, the Kairos World Model is transforming world models from laboratory demonstrations into industrial infrastructure that supports real-world robot operation.
Company Profile
Ace Robotics focuses on foundational research and development for embodied intelligence world models. Bringing together internationally recognized scholars and doctoral research partners, the company pioneered a complete technical loop that integrates ACE ambient data collection, world models, and generalization modules. It delivers robot brain solutions across industrial applications, urban governance, retail, culture, and tourism.
Media Contact: ACE Robotics (Shanghai) Co., Ltd. Mr. Miao Zhen Tel: +86 186 1602 4619 Official Page: https://www.linkedin.com/company/acerobotics/
Legal Disclaimer:
EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
