Xiaomi-Robotics-1: Pre-Training Meets Real-Home Data
20 Jul 2026
Xiaomi has introduced Xiaomi-Robotics-1, a robot learning model that pairs large-scale embodiment-free pre-training with a comparatively modest amount of real-robot data collected in actual homes. The model is shown performing household tasks such as packing and laundry loading after this two-stage training process.
What Xiaomi Built
The system's training pipeline unfolds in two stages:
- Pre-training: Uses 100,000 hours of embodiment-free (UMI) trajectories spanning more than 1,700 scenarios.
- Post-training: Uses over 7,200 hours of real-robot data collected in real homes.
According to the source, Xiaomi-Robotics-1 can learn new tasks — including phone packing, printer refilling, laundry loading, and box packing — from just a few hours of real-robot demonstrations per task. The company also reports state-of-the-art results on four mainstream simulation benchmarks, though none are named specifically in the available material.
The Numbers Behind the Claims
Evaluation results show a clear tradeoff between demonstration time and reliability:
- 75% overall success rate with an average of under 10 hours of demonstrations per task.
- 85% overall success rate with an average of under 40 hours of demonstrations per task.
The source also states that success rates keep improving as the model consumes more data or scales up during pre-training, with no signs of saturation — though this claim, like the benchmark comparisons, is self-reported and not independently verified.
What's Missing
Several details are notably absent from the available information: there's no announcement or deployment timeline, no specifics on the hardware platform the model runs on, no named competitor comparisons, no cost or compute figures, and no breakdown of how many homes or what demographics the real-world data came from. Failure modes behind the 75%/85% success rates also aren't detailed.
Why Founders Should Care
For founders building in robotics or embodied AI, this report suggests a few probable takeaways:
- The pre-training/post-training split may indicate a viable strategy for cutting real-world data collection costs — a persistent bottleneck in robotics startups.
- If the reported non-saturating scaling trend holds, it could suggest that continued investment in data and model scale may keep paying off, which would be relevant for teams building on similar embodiment-free pre-training approaches. This is currently based solely on the source's own claims and lacks independent verification.
- The ability to learn tasks from just a few hours of demonstration data hints at potentially lower barriers to customizing robotic systems for new use cases — though founders should likely budget for tens of hours of real-robot data per task if aiming for the higher end of reported reliability (85%).
The Caveats
The 75-85% success rates, while notable, mean the system may still fail in a meaningful share of task attempts — a consideration for anyone evaluating readiness for consumer or commercial deployment. Additionally, reliance on real-home data collection raises unaddressed questions around privacy, safety, and data provenance that the source does not cover. Founders tracking this space should treat the state-of-the-art and non-saturation claims as directional signals rather than confirmed benchmarks until independently verified.