Xiaomi-Robotics-1: Scaling Robot Policies with 100K Hours of Data
Xiaomi just dropped a robot foundation model that breaks the data bottleneck. Xiaomi-Robotics-1 (XR-1) is pre-trained on 100,000 hours of embodiment-free (UMI) trajectories spanning over 1,700 scenarios. The key insight: you don't need robot-specific data for pre-training. Instead, you use human manipulation videos (UMI data) and then align to real robots with a small post-training set.
Scaling Laws for Robot Policies
The team found that pre-training shows clean scaling behavior. As data and model size grow, validation action error steadily decreases. After post-training, real-robot success rate rises predictably with pre-training data volume and model size. No saturation seen yet. This mirrors language model scaling but for physical action.
Data Pipeline
Pre-training uses 100K hours of UMI trajectories across household, commercial, industrial, and outdoor environments. The auto-labeling pipeline splits long videos into fixed-length clips and uses a vision-language model (VLM) to describe state transitions of grippers and interacting objects. This produces a large corpus of manipulation trajectories with precise language annotations.
Post-training uses cross-embodiment datasets: 7,200 hours of in-house real-robot data collected in real homes (tasks like tidying a sofa, sorting a shoe cabinet, putting away kitchenware), filtered open-source robot data, and high-quality UMI data manually annotated with temporal segments and instruction prompts.
Training Paradigm
Two stages:
- Pre-training: Learns general representations for action generation from large-scale UMI data. Model learns to generate actions that drive scene state transitions described by language.
- Post-training: Aligns to real robot embodiments and instruction following. Embodiment alignment uses cross-embodiment real-robot data. Instruction alignment shifts from "generate actions given state transition description" to "understand natural-language instruction and execute."
Real-World Performance
After post-training, XR-1 works out-of-the-box in unseen environments with unseen objects. The team evaluated in real homes and found that stronger pre-trained models yield better real-robot performance.
Efficient Adaptation
XR-1 adapts to new tasks from minimal data. With under 10 hours of demonstrations per task (phone packing, printer refilling, laundry loading, box packing), it achieves 75% overall success rate, nearly doubling the π0.5 baseline (40%). With under 40 hours, it reaches 85%.
| Task | <10h/task (XR-1) | <10h/task (π0.5) | <40h/task (XR-1) | <40h/task (π0.5) |
|---|---|---|---|---|
| Phone Packing | 70 | 30 | 80 | 40 |
| Printer Refilling | 70 | 20 | 60 | 20 |
| Laundry Loading | 80 | 40 | 100 | 50 |
| Box Packing | 80 | 70 | 100 | 100 |
| Overall | 75 | 40 | 85 | 53 |
Simulation Benchmarks
XR-1 achieves state-of-the-art on four mainstream simulation benchmarks:
| Benchmark | XR-1 | 2nd Best | Rel. Gain |
|---|---|---|---|
| RoboCasa | 74.5 | 72.6 | +2.6% |
| RoboCasa365 | 57.4 | 46.6 | +23.2% |
| VLABench | 59.1 | 53.2 | +11.1% |
| RoboDojo | 13.9 | 8.80 | +58.3% |
What This Means
XR-1 demonstrates a practical path for scaling robot foundation models. The embodiment-free pre-training breaks the data bottleneck. The scaling behavior is predictable and transfers to real robots. This could accelerate development of general-purpose robots.
Next Steps
Xiaomi released the paper on arXiv (arXiv:2607.15330). The model itself isn't open-source yet, but the methodology is fully described. If you're building robot policy models, study their auto-labeling pipeline and two-stage training. The 100K-hour pre-training dataset is a significant resource — consider how to adapt it to your own embodiments.
Check the source article for uncut footage of luggage packing.
