China’s embodied-AI field is re-examining its own advantages. The supply chain and data edges it once took for granted are being questioned: many suppliers does not mean fit for embodied use, and large data volume does not guarantee intelligence. As one company puts it, distribution matters more than raw volume. What China needs next is to let field problems drive original architecture and build its own “root technology”.
On 14 September the “Beyond Data” forum on brain-inspired embodied cognition opened in Shanghai as a sub-forum of the 2026 Pujiang Innovation Forum. EBKernel, the organiser, argues plainly for “clear brain-inspired mechanisms instead of blind computation”. Starting from the world’s first brain-inspired cognitive world model, Cog-WM 1.0, the company wants robots to understand the physical world through memory, prediction and active exploration.
Teaching a robot to remember
A robot that reaches a shop in a mall completes the navigation task. But a mall that will use the robot long term cares about what comes after: does it remember the route, can it find an object it saw earlier, does it need to relearn the path next time? People do not relearn the world from zero on each trip. EBKernel wants to give robots the ability to carry one experience into the next action.
Cog-WM 1.0 ships two branches, Nav and Manip. Nav is the first validated entry: traditional navigation reads a map, Cog-WM keeps a memory. It organises spatial structure separately from objects and events, stores their relations in a Graph-Voxel global spatio-temporal memory, and updates what is worth remembering when new observations diverge. In hidden space it predicts action outcomes and focuses computation on decision-relevant abstractions.
In a live demo the robot had no preset 3D map yet found targets in unfamiliar spaces; a plant glimpsed by chance became a retrievable memory, and the robot could return to it and describe what sat beside it. On the 50 per cent HM3D ObjectNav subset the Nav branch reached 86.89 per cent success using 145 scenes, 577 trajectories and about 15.4 hours of training data.
The manipulation branch and the bigger claim
The Manip branch learns from consecutive actions. Through multi-timescale prediction it trains on near-term action outcomes and later task states; through value-guided experience learning it judges how much each action segment advances the task. On LIBERO, LIBERO-Plus and RoboTwin 2.0 Hard it reached 98.2 per cent, 84.6 per cent and 43.2 per cent respectively, lifting RoboTwin 2.0 Hard by 6 percentage points, a 16.2 per cent relative gain, over the pi0.5 baseline.
EBKernel’s CEO Zhu Senhua frames the route as a “root technology” whose logic is industrial security and value distribution. Whoever defines the base-layer paradigm takes the thickest slice of the value chain. In mobile, ARM took the chip architecture, Google and Apple took the OS; the application layer prospered but the profit was set at the base. If the base paradigm of embodied AI is defined overseas, China’s supply-chain and scenario edges are built on someone else’s land.
The company is not alone in the direction. LeCun’s AMI Labs, Karl Friston’s active-inference framework, Verses, Google DeepMind and Meta are all investing here. EBKernel borrows the JEPA idea of prediction in abstract hidden space and adds spatial memory, surprise-driven updates, multi-timescale prediction and value-guided learning. Zhu’s standard for a true third pole is blunt: it must produce real, repeatable commercial results in the field, and more teams must enter and form an ecology. “If only EBKernel walks this road, that is an outlier. One pole cannot be one company.”
China’s embodied AI today is “three strong, one weak”: hardware supply chain, scenarios and engineering iteration are strong, while the original embodied brain is relatively thin. The moment to build that brain has arrived. Hot markets favour speed; root technology needs patience, and patience must keep producing results.
Editor’s note: This is an adapted translation of the original OFweek Robotics report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-09/ART-8321205-8610-30703198.html.