Who Can Crack the Warehouse: China's First Big Test for Embodied AI
Short videos show waves of humanoid robots picking packages in warehouses. This reflects how logistics warehousing is becoming the first major battleground for embodied AI in China.
Yet warehouses don't care about demos; they care about efficiency. A warehouse is a complex system where any weak link drags down the whole process—making robots faster or smarter alone cannot solve the problem.
Atomix, a logistics robotics company with a decade of business experience, believes the answer lies in system-level capability.
In a warehouse, trucks dock at the platform, pallet AMRs move goods into dense storage, and four-way shuttles switch directions between racks. When orders arrive, the system calculates inventory locations, order combinations, equipment loads, and downstream rhythms simultaneously before deciding which pallet to pick first, which device to assign, and where to hand off.
Tasks with clear rules and repetitive motions are handled by specialized equipment. Items that are hard to identify or grasp, such as soft packages or irregular shapes, go to robots with vision and manipulation capabilities. Exceptions beyond robotic limits are passed to human workers.
Orchestrating everything is a multi-agent layered robotic operation system that connects robots, equipment, and people into a network, breaking an order into hundreds of actions and deciding who does what, when to hand off, and how to recover from errors.
In 2018, the Atomix team—then within Megvii—won the Uniqlo Shanghai automated warehouse project after five rounds of selection. Implementation was tough: only 30% of technical goals were met at launch, and the team spent half a year rewriting code. This experience won Uniqlo's global follow-up orders and cemented two beliefs: warehouse efficiency depends on overall scheduling, not individual machine performance; and customers buy actual benefits, not robots.
Based on this logic, the team developed a 4+1 product framework: four types of standard hardware—pallet four-way shuttles, tote four-way shuttles, pallet AMRs, and tote AMRs—plus a unified software platform, Atomixer. The platform uses over 50 proprietary algorithms to schedule equipment, containers, and orders.
In warehousing, the more devices, the exponentially higher the probability of conflicts, deadlocks, and congestion. Scheduling scale is therefore a hard metric. Atomix is the world's first company to deploy over 80 pallet four-way shuttles in a single system, and the company with the most super-large four-way shuttle projects (over 50 units) globally.
By the end of 2025, Atomix had annual revenue approaching 1 billion yuan, covering over 20 countries and nearly 100 brands. It ships nearly 10,000 robots annually, ranks second globally in pallet four-way shuttle sales, and has served over 500 projects.
On June 5, 2026, Yuanli Lingji, an embodied AI company also founded by Tang Wenbin, announced a merger with Atomix through equity acquisition. One side brings scaled performance and operational data in a key embodied AI scenario; the other brings DM0.5, the world's first embodied-native large model. The merger gives the model a scenario and the scenario a model.
Tang Wenbin graduated from Tsinghua University's Yao Class and is a co-founder of Megvii. In 2016, he initiated robotics and warehousing business within Megvii, starting from scheduling and launching the robot cluster scheduling hub HETU in 2019. The path—scenario first, then devices, then models—is almost the reverse of the industry's mainstream approach, but it means the model grows inside the warehouse from day one.
Warehouse intelligence must overcome three hurdles. First, legacy: warehousing has established order, systems, and standards, so new solutions must accommodate existing assets and upgrade gradually without disruption. Second, elasticity: during Double 11 in 2025, daily parcel peaks reached 777 million, about 1.4 times normal; systems must add or remove devices, adjust routes, and bring in human workers for exceptions and peaks. Third, long tail: tasks like shelving, breaking bulk, soft package grasping, and irregular item sorting still rely heavily on manual labor, and needs vary greatly across warehouses.
Tang has repeatedly said embodied AI is far from its ChatGPT moment. He sets strict acceptance criteria: continuous use means running at least 10 hours a day for two consecutive months, with 100 units as a starting point and 1,000 units as solid proof.