Breaking the Data Barrier: Scaling Robot Policy Models with Embodiment-Free Pre-training
Xiaomi-Robotics-1 is a robot foundation model trained on over 100K hours of real-world manipulation trajectories. It is a Vision-Language-Action (VLA) model engineered for out-of-the-box mobile manipulation in unseen environments and efficient adaptation to new tasks.
XR-1 follows a two-stage training paradigm inspired by large language models — pre-training for breadth, followed by post-training for alignment. It showcases that pre-training scaling behavior reliably transfers through post-training to real-world robot performance, with no signs of saturation.
XR-1 couples a pre-trained VLM (Qwen3-VL) with a Diffusion-Transformer (DiT) via a Mixture-of-Transformers (MoT) — the DiT matches the VLM in layer count but uses a smaller hidden size for faster inference.
- 🧠 Strong Generalization: Pre-trained on 100K+ hours of embodiment-free UMI trajectories spanning 1,700+ scenarios across household, commercial, industrial, and outdoor environments, enabling robust handling of complex, unseen tasks.
- 🚀 Real-Time Ready: Optimized with asynchronous execution to minimize inference latency.
- 🛠️ Flexible Deployment: Built on the Hugging Face
transformersecosystem (pinned to4.57.1) and optimized for consumer GPUs. - ⚡ Data-Efficient Adaptation: Learns entirely new complex tasks from just a few hours of demonstrations per task.
Stage 1 — Pre-training (breadth). Uses embodiment-free UMI data to learn generalizable representations for action generation. A VLM-powered auto-labeling pipeline automatically segments long trajectories into fixed-length clips and generates descriptions of the corresponding state transitions. Trained on these annotated clips, the model learns to generate actions that bring each scene from its current state to the described target state. By making large-scale annotation practical, this approach eliminates the need for prohibitively expensive manual labeling and breaks the robot-data scarcity bottleneck that has historically constrained policy-model scaling.
Stage 2 — Post-training (alignment). Aligns the pre-trained model along two axes with cross-embodiment data, including in-house real-robot data, filtered open-sourced robot data, and high-quality manually-annotated UMI data:
- Embodiment alignment — maps general action-generation ability onto actual robots via cross-embodiment data.
- Instruction alignment — shifts from acting on scene-transition descriptions to understanding natural-language instructions and executing them directly.
After post-training, XR-1 can perform mobile manipulation in unseen environments with unseen objects, out of the box. The model, pre-trained on over 100K hours of UMI trajectories and subsequently post-trained on over 10K hours of cross-embodiment data, has been open-sourced on Hugging Face.
- [2026-08-03] 💻 Released code and checkpoints for post-training, inference, and benchmark evaluation.
- [2026-07-16] 📄 Released technical report (arXiv:2607.15330).
We evaluate Xiaomi-Robotics-1 on four standard simulation benchmarks: RoboCasa, RoboCasa365, VLABench, and RoboDojo. XR-1 achieves state-of-the-art results on all four. As of July 15, 2026, it ranks first on both the RoboCasa365 and RoboDojo leaderboards.
| Benchmark | XR-1 | 2nd Best | Relative Gain |
|---|---|---|---|
| RoboCasa | 74.5% | 72.6% | +2.6% |
| RoboCasa365 | 57.4% | 46.6% | +23.2% |
| VLABench | 59.1% | 53.2% | +11.1% |
| RoboDojo | 13.93% | 8.80% | +58.3% |
For each setting, we provide the corresponding fine-tuned checkpoint and a guide for running the evaluation:
| Benchmark | Checkpoint | Evaluation Guide |
|---|---|---|
| RoboCasa | Xiaomi-Robotics-1-RoboCasa | eval_robocasa/README.md |
| RoboCasa365 | Xiaomi-Robotics-1-RoboCasa365 | eval_robocasa365/README.md |
| VLABench | Xiaomi-Robotics-1-VLABench | eval_vlabench/README.md |
| RoboDojo | — | official docs |
XR-1 adapts to new complex tasks from minimal demonstration data, substantially outperforming the π0.5 baseline:
| Task | XR-1 (< 10h/task on avg.) | π0.5 (< 10h/task on avg.) | XR-1 (< 40h/task on avg.) | π0.5 (< 40h/task on avg.) |
|---|---|---|---|---|
| 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% |
Each cell shows success rate, higher is better. XR-1 = Xiaomi-Robotics-1.
Want to fine-tune XR-1 on your own data? The post-training and deployment code is available in xr1/. See xr1/README.md for details on environment setup, checkpoint conversion, data formats, training, and deployment.
For discussions, scan the QR code below to join our WeChat group.
If you find this project useful, please consider citing:
@article{team2026xiaomi,
title={Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories},
author={Team, Xiaomi Robotics and Guo, Jun and Jin, Piaopiao and Li, Jason and Li, Peiyan and Li, Yingyan and Liu, Futeng and Peng, Wanli and Qin, Optimus and Su, Yifei and others},
journal={arXiv preprint arXiv:2607.15330},
year={2026}
}This project is licensed under the Apache License 2.0.