Kunteng says warehouse robots clear PMF validation as orders approach 5,000 units

China's 2026 Double 11 shopping festival has ended, but the warehouses it filled are only now hitting peak workload. Picking, packing, weighing, dimensioning and parcel feeding are the hardest roles to staff, and this year Kunteng's embodied robots worked inside Cainiao's supply chain during the peak for the first time, running against real holiday order pressure rather than a demonstration line.

In November the company said it had completed product-market fit validation across weighing and dimensioning, sorting and feeding, and fixed picking, with deployments in warehouses run by several leading e-commerce platforms and logistics firms. Those deployments have been replicated across sites and produced repeat purchases, and Kunteng says it now holds batch orders for nearly 5,000 robots, moving into volume customer acquisition and delivery.

Logistics warehouses are where physical AI is finding traction first, because the tasks are dense and the feedback is fast. A conveyor moves at roughly 0.3 metres per second, every parcel differs in shape, wrapping and orientation, and weighing, dimensioning, label swapping and feeding all have to hand off to the conveyor and the warehouse system while keeping pace. A failed motion does not stop the line, so the failure becomes a training sample instead of a shutdown.

Cross-border logistics was the first beachhead. Kunteng says talks with a leading cross-border e-commerce platform began in July, pilots ran in eastern and southern China, robots entered a warehouse in September and volume purchase contracts were signed in October. Its general-purpose dual-arm robot has completed deployment validation for weighing and feeding posts, reaching more than 1,200 parcels per hour with a 99% success rate.

The numbers rest on model and systems work. Kunteng trains visual perception and action policy jointly so a single policy can absorb different packaging, materials, poses and occlusions instead of relying on per-SKU programming, and it uses an asynchronous input model in which cameras, tactile sensing and robot state enter at their own cadence rather than being force-aligned. Field failures, human takeovers and recoveries feed a data loop, with post-training and regression testing cycling in days, and a Fleet Learning effort is meant to push fixes across the whole robot fleet.

Customer demand is equally concrete. In September Kunteng signed a strategic partnership with Kejie Group to build logistics embodied-AI solutions, and the first pilot is in implementation: an initial positive-picking idea was reworked on site into night-shift put-away for apparel returns, relieving pressure during concentrated return waves. Pilots with SF Express are under discussion for sorting-centre feeding and mobile picking, with AGV scheduling and parcel-cage loading as next candidates.

Overseas, Kunteng's robots have entered Cainiao's Hong Kong eHub for demonstrations to global media and cross-border customers; a North American subsidiary has been set up and is signing agreements with large retailers; in Korea the company plans a joint venture and co-developed products with a listed robotics firm; and in Europe and Latin America it has signed local partner and pilot agreements. Founder and CEO Li Qiang was previously CTO of Cainiao Group, and the team has run L4 delivery vehicles at thousand-unit scale.

The next checkpoint is product cadence: Kunteng plans to launch a robot with an entirely new configuration at the end of December 2026, designed around the constraints of real workstations and aimed at applicability, continuous operation and cost. PMF validation and a five-thousand-unit order book are only the entry ticket; whether cross-warehouse replication, delivery operations and unit economics can move together is what decides whether scale is real. Disclosure: this report is based on material supplied by Kunteng and republished by QbitAI with authorisation; all figures are the company's own.