Xiaomi's Humanoid Robots Hit 98% on the Factory Floor — But Full Replacement Is Still Years Away

In the initial March trial, the robot worked autonomously for three continuous hours, performing simultaneous dual-sided nut installation with a 90.2% success rate while matching the production line's required 76-second cycle time . That trial earned the robots the nickname "interns" from Xiaomi President Lu Weibing, who told CNBC that two humanoid robots could complete over 90% of assigned work during a three-hour shift, handling tasks such as nut installation and material movement .

Expansion to More Complex Tasks

Beyond nut installation, Xiaomi has expanded the robots' responsibilities. The company introduced two new tasks: center console side panel sorting and parts bin folding and recycling, both achieving success rates of 90% . In April 2026, Xiaomi debuted a new full-size humanoid robot at its Investor Day — the result of four years of development — that had already been tested in factories with over a 90% success rate in complex installation tasks .

Xiaomi-Robotics-U0: Open-Source AI Model Boosts Robot Performance

On July 15, 2026, Xiaomi released and fully open-sourced Xiaomi-Robotics-U0, a 38-billion-parameter multimodal autoregressive world foundation model for embodied intelligence . Key features include:

• Unified framework: It handles four categories of embodied generation tasks — text-to-image, image editing, embodied scene generation, and embodied video generation — in a single autoregressive framework .
• Architecture: Built on EMU3.5 and Qwen-3-32B, it uses a "next-token prediction" architecture to simulate physical world interactions .
• Speed: A FlashAR+ inference acceleration scheme improves generation efficiency by approximately 83× over the original autoregressive paradigm .
• Benchmark performance: It placed first on the World Arena benchmark for embodied video generation and beat GPT-Image-2.0 in human evaluation for embodied scene generation .

Real-world impact: When a robot policy (named π₀.₅) was trained with synthetic data generated by U0, its real-world out-of-distribution (OOD) success rate jumped from 36.9% to 63.2% — a 26.3 percentage point improvement .

Published Iteration Curve

This represents a 7.8 percentage point gain over four months. CEO Lei Jun has stated the goal is to deploy "a large number" of humanoid robots across Xiaomi's production facilities within the next five years .

Remaining Gaps: From "Interns" to Full-Time Operators

Despite rapid progress, the robots face significant hurdles before they can operate as full-time assembly line workers.

1. Cycle time: The 76-second cycle time was matched in March, but automotive lines typically demand faster paces. Whether the robots can maintain that speed across a full multi-hour shift without degradation is unconfirmed .
2. Uptime and reliability: Trials have lasted only three hours. Full-time operators work 8+ hour shifts with near-zero unscheduled downtime. Xiaomi has not published longer-duration reliability data .
3. Defect costs: At a 98% success rate, the 2% failure rate translates to real costs in automotive assembly — rework, scrap, and potential downstream quality issues. Human workers at 99%+ still have an edge, and safety-critical tasks often require defect rates measured in parts per million .
4. Task breadth: Nut installation is a single, relatively constrained task. Scaling to the dozens of varied, dexterous assembly operations required across a full vehicle line is a substantially harder problem. The expansion to flexible workpiece handling and material movement is promising but still limited .

In short, Xiaomi's humanoid robots have moved from pure lab experiments to real-world factory trials with measurable, improving results. But as company president Lu Weibing put it, they are still interns — not yet ready for a full-time job.