Xiaomi Humanoid Hits 90% Success in Factory Trials
Xiaomi's humanoid robot is now undergoing real-world testing in a car factory, working for over three hours straight. The robot is autonomously installing self-tapping nuts with a greater than 90% success rate, matching the 76-second pace of the production line. Xiaomi is now targeting large-scale deployment within five years.
The robot in the trial, initially unveiled in August 2022 as CyberOne, stands 177cm tall and weighs 52kg. Its early demonstrations focused on bipedal motion and AI-driven emotional interaction, with the ability to recognize 45 human emotions and 85 environmental sounds. The hardware features 21 degrees of freedom with a joint response time of 0.5ms and can carry up to 1.5kg in a single hand. Underpinning the robot's factory performance is a Vision-Language-Action (VLA) foundation model named Xiaomi-Robotics-0. This end-to-end, data-driven approach allows the robot to learn from interactions in the physical world, reducing reliance on teleoperation data and enabling it to adapt more quickly to downstream tasks. To handle the physical complexity of the task, which includes inconsistent nut orientation and magnetic interference from positioning pins, Xiaomi employs a second specialized model called TacRefineNet. This system fuses data from vision, touch, and joint proprioception, using tactile feedback to fine-tune grasping and improve the stability and robustness of manipulation. The robot maintains balance on the factory floor using a hybrid motion control architecture. An optimization-based controller handles prioritized safety constraints with a solve time under 1ms, while a separate reinforcement learning controller, trained on a massive parallel simulation platform, manages recovery from extreme disturbances, allowing for zero-shot deployment from simulation to the real world. This factory trial places Xiaomi in direct competition with other major players deploying humanoids in industrial settings. Tesla's Optimus is a key rival, and other automotive firms like Hyundai (backing Boston Dynamics) and XPeng are also developing their own robots. Meanwhile, startups like Figure AI and Agility Robotics are also targeting logistics and manufacturing applications. The move from a lab setting to a live production line is a significant step, as noted by CEO Lei Jun. While failures are acceptable during R&D, the consistency and reliability required to meet the cycle times of an automotive assembly line present a much greater engineering challenge.
