Xiaomi Deploys Humanoid Robots on Its EV Assembly Line, Achieving 90 Percent Success Rate in Autonomous Nut Installation
Xiaomi's humanoid robots completed three hours of autonomous assembly work at its EV plant, hitting a 90.2 percent success rate and meeting the 76-second production cycle, as CEO Lei Jun pledges large-scale factory deployment within five years.
Overview
Xiaomi has begun deploying self-developed humanoid robots at its electric vehicle assembly plant, where the machines operated autonomously for three continuous hours performing precision tasks on a live production line. The company announced on March 2 that the robots achieved a 90.2 percent success rate for simultaneous installation of self-tapping nuts on both sides of a workstation, meeting the line’s 76-second-per-unit cycle time, according to the South China Morning Post.
CEO Lei Jun described the trial as marking “the first step towards stable application” of humanoid robots in intelligent manufacturing, and pledged that Xiaomi would deploy “a large number” of the machines across its factories within the next five years.
What We Know
The robots were tested in the die-casting workshop of Xiaomi’s EV plant, where they performed multiple tasks including picking self-tapping nuts from automatic feeding devices, placing them onto positioning fixtures, and coordinating with slide conveyors to complete automated tightening of floor components after integrated die-casting. Additional demonstrated tasks included removing protective films, installing vehicle badges, and moving material boxes, as detailed by the South China Morning Post.
The assembly work presents significant engineering challenges. The robots must handle precise alignment under magnetic force interference, manage varied gripping postures, and maintain sub-millimeter accuracy across repetitive cycles. Xiaomi’s solution relies on a hybrid approach that combines an optimization-based controller, which updates in under one millisecond, with reinforcement learning trained on hundreds of millions of simulated disturbances, as reported by Interesting Engineering.
Underpinning the system is Xiaomi-Robotics-0, a 4.7-billion-parameter open-source Vision-Language-Action (VLA) foundation model that the company released in February. The model uses a Mixture-of-Transformers architecture and integrates multimodal perception spanning vision, tactile feedback, and joint proprioception. According to its technical paper on arXiv, the model achieved state-of-the-art results across the LIBERO, CALVIN, and SimplerEnv simulation benchmarks, and is designed to generate high-frequency, continuous action outputs in real time.
Xiaomi also deployed a tactile fine-tuning model called TacRefineNet, which relies primarily on touch sensing rather than vision for the final stages of precision assembly. This allows the robot to handle alignment corrections even when visual occlusion would defeat camera-based approaches.
The Broader Race
Xiaomi’s factory trial arrives amid an intensifying global contest to bring humanoid robots from laboratories to production floors. Chinese domestic firms are shipping more units and iterating faster than their American counterparts in a still-nascent market, as TechCrunch reported.
Xpeng plans to break ground on a 110,000-square-meter humanoid robot factory in Guangzhou this quarter and targets mass production of its IRON robot by late 2026, with a stated goal of one million units annually by 2030. Unitree shipped over 5,500 humanoid robots in 2025 and plans to reach 20,000 this year. Meanwhile, Tesla deployed over 1,000 Optimus Gen 3 units across its factories in January 2026, though CEO Elon Musk clarified on an earnings call that the machines are primarily in a data-gathering phase rather than performing production work.
Boston Dynamics confirmed that its electric Atlas robot will begin commercial deployment at Hyundai’s Metaplant in Georgia during 2026, featuring 56 degrees of freedom and a 110-pound lifting capacity.
What We Don’t Know
Xiaomi has not disclosed the total number of humanoid robots currently deployed, the unit cost per robot, or the specific factory location beyond its EV manufacturing operations. The company also has not detailed how performance metrics such as mean time between failures compare to existing industrial automation equipment already on its lines.
The 90.2 percent success rate, while notable for a humanoid performing dexterous assembly, still implies roughly one failure in every ten attempts. Whether this translates to acceptable defect rates at full production scale, or whether it requires human operators to remain on standby for error correction, is not yet clear.
Whether humanoid form factors offer meaningful advantages over purpose-built robotic arms for structured factory tasks also remains an open question. Humanoid robots are substantially more expensive and mechanically complex than single-purpose automation, and the business case for deploying them in environments already designed around conventional robotics has yet to be proven at scale.
Analysis
Lei Jun framed the deployment in terms far broader than a single factory trial. He predicted that humanoid robots would create “a new trillion-yuan industrial market” and argued that household robots would eventually become an even larger market than factory applications, according to Interesting Engineering. He also emphasized that “no single company can build this market alone,” advocating for partnerships and shared engineering platforms across the industry.
Xiaomi’s decision to open-source its VLA foundation model signals a strategy familiar from its smartphone playbook: commoditize the software layer to accelerate ecosystem development while retaining advantages in hardware integration and manufacturing scale. The company already operates what it calls a “dark factory” for smartphone production and has been expanding engineering teams focused on AI, robotics, and smart EV systems.
The five-year deployment window also reflects a practical acknowledgment that humanoid robotics remains far from the reliability standards required for continuous factory operation. Xiaomi’s own data shows that key performance indicators, including mean time between failures and single-task success rates, are still “steadily improving” rather than at production-ready levels. The gap between a three-hour demonstration and 24/7 operation across multiple workstations remains substantial.