EBKernel Releases Cog-WM 1.0, a Brain-Inspired Cognitive World Model for Robots

The model is built on what EBKernel describes as a systematic set of brain-inspired neural mechanisms combined with the Joint Embedding Predictive Architecture, or JEPA. Rather than predicting pixel-level changes, it predicts in a latent space how task-relevant space and state will evolve, and uses those predictions to drive action. The company lists three mechanisms: perception encoding with goal modulation, which rebuilds the world at the level of abstract representations rather than pixels; separation of the content and structure of spatiotemporal memory, so spatial structure learned in one room can be reused in another and learning is focused on surprising observations; and multi-level latent prediction across time and space scales, which the company says supports long-horizon exploration. Navigation and manipulation use different implementations of this shared framework, with navigation building explicit spatial memory to predict exploration directions and manipulation learning state changes over multiple time scales from execution experience.

On navigation, EBKernel's algorithm team reported that on a subset of the HM3D-ObjectNav benchmark, Cog-WM Nav 1.0 raised the success rate to 86.89 percent from 78.50 percent, an 8.39 percentage point gain, or a 10.69 percent relative improvement, over the BSC-Nav baseline published in Nature Communications (Ruan et al., 2026). Path efficiency, measured by SPL, rose to 48.35 from 47.70. The company says the system requires no pre-built map prior.

For manipulation, Cog-WM Manip 1.0 outperformed the pi-0.5 baseline, which was pretrained on massive datasets, by up to 16 percent or more across three mainstream embodied manipulation benchmarks, one of which was RoboTwin 2.0 Hard, described by the company as among the most difficult. An ablation study on LIBERO-Plus reported a stepwise rise in success rate: 80.0 percent for a pure policy, 81.6 percent with local prediction added, 82.0 percent with multi-horizon prediction, and 84.6 percent for the full method. EBKernel said the model remained reliable when lighting, viewpoint and background texture changed. The release did not list the individual success rates for the three manipulation benchmarks, and the figures were disclosed by the company's algorithm team.

Methods behind both branches have been published as arXiv papers, the company said. Cog-WM Nav 1.0 has been deployed on quadruped robots at customer sites for inspection and patrol, and its navigation capabilities, including autonomous navigation and path planning without a pre-built map, spatiotemporal memory retrieval, spatial-relation question answering and object search, have been validated on several wheeled humanoid platforms. Manipulation has been validated on a wheeled humanoid robot.

Zhu Senhua, founder and CEO of EBKernel, said embodied intelligence is at a critical stage of moving from the laboratory to industrial deployment and that low data dependence and high generalization are the core proposition of that process. He said Cog-WM 1.0 offers preliminary validation that drawing systematically on the brain's functional neurocognitive mechanisms can raise the ceiling of algorithmic capability, and that scaling data alone is not enough; the structure and architecture of models should also be improved. He described brain-inspired embodied intelligence as a road few take, because it is a road few can take, but one he called strategically correct.

EBKernel was founded in June 2025 and says its team comes from Huawei, Lenovo, Megvii, Geek+, Sun Yat-sen University, the University of Pennsylvania, the Chinese Academy of Sciences and Tsinghua University, with more than a decade of cross-disciplinary AI and neuroscience research, 15 years of robotics supply-chain coordination and experience deploying robots at the 10,000-unit scale. The forum where the model was released is billed as the first themed forum on brain-inspired embodied intelligence, bringing together researchers from Shanghai, Beijing and Shenzhen and from the China Brain Project's South Brain and North Brain programs, along with chip makers, algorithm developers and investors. Brain-inspired intelligence is listed among the future industries in China's 15th Five-Year Plan, according to the release.

Editor's Summary

EBKernel released Cog-WM 1.0, a brain-inspired latent-space world model, on Sept. 14, claiming leading results without pre-built maps or large-scale pretraining data. The company reported an 86.89 percent navigation success rate on the HM3D-ObjectNav subset and manipulation results above the pi-0.5 baseline, with deployment already under way on quadruped and wheeled humanoid robots. The figures come from the company and are not yet backed by independent evaluation.