Xiaomi‑Robotics‑1 scaling study: embodiment‑free pre‑training breaks the robot data barrier

Xiaomi‑Robotics‑1 scaling study: embodiment‑free pre‑training breaks the robot data barrier

Xiaomi‑Robotics‑1 proves that large‑scale, embodiment‑free pre‑training can break the data bottleneck in robot policy learning

Takeaway: Xiaomi‑Robotics‑1 demonstrates clean scaling of robot foundation models by pre‑training on 100 k hours of unlabeled real‑world video (UMI) and then aligning with a modest amount of real‑robot data, yielding steady improvements in real‑robot success rates and state‑of‑the‑art performance on simulation benchmarks.

Scaling laws apply to robotics when data is abundant

Conclusion: When the amount of pre‑training data and model size increase, validation action error drops consistently, and the same trend carries over to real‑robot success rates after post‑training.

Xiaomi‑Robotics‑1 follows the two‑stage paradigm used for large language models:

  1. Pre‑training on 100 000 hours of embodiment‑free (UMI) trajectories covering 1 700+ scenarios. An automatic pipeline uses a strong vision‑language model to annotate fixed‑length video clips with language descriptions of scene‑state transitions.
  2. Post‑training on 7 200 hours of in‑house robot data plus filtered open‑source datasets, aligning the learned action‑generation ability with actual robot embodiments and natural‑language instructions.

The authors report a clean scaling curve for pre‑training (see Figure 3 in the original release) and a monotonic increase in real‑robot success rate as the pre‑trained model grows (Figure 4). No saturation is observed within the explored data and model size range.

Data strategy eliminates the traditional robotics data bottleneck

Conclusion: Using embodiment‑free video as a pre‑training source provides orders of magnitude more data than conventional robot‑centric collections, while a lightweight post‑training stage transfers this knowledge to physical robots.

  • Pre‑training data: 100 k h of raw video from households, commercial spaces, industrial sites, and outdoors. The auto‑labeling pipeline generates language captions for each clip, enabling the model to learn a mapping from language to manipulation actions without any robot‑specific supervision.
  • Post‑training data: 7 200 h of real‑robot trajectories (e.g., sofa tidying, shoe‑cabinet sorting, kitchenware storage) plus high‑quality UMI segments with manually written instruction prompts.

This split mirrors the “breadth‑first” approach of LLMs: first acquire a general world model, then specialize to a target embodiment.

Real‑world performance scales predictably with model size

Conclusion: Stronger pre‑trained models consistently achieve higher success rates on unseen real‑world tasks after post‑training.

The authors evaluated Xiaomi‑Robotics‑1 in novel environments with unseen object instances. Results show a steady rise in success rate as the pre‑training data volume or model parameters increase, confirming that scaling benefits transfer from simulation‑style validation to physical execution.

"The answer is yes. As we increase the amount of pre‑training data and model size, real‑robot success rate rises steadily and predictably. That is, a stronger pre‑trained model yields better real‑robot performance."

Efficient adaptation to new tasks with minimal demonstrations

Conclusion: Fine‑tuning Xiaomi‑Robotics‑1 on a few hours of task‑specific data yields high success rates, demonstrating data‑efficient learning.

When adapted to four downstream tasks (phone packing, printer refilling, laundry loading, box packing), the model achieved:

  • ~75 % success with <10 h of demonstrations per task.
  • ~85 % success with <40 h per task.

These figures nearly double the baseline (≈40 %). The results illustrate that a large, generalist foundation model can be specialized quickly, reducing the cost of collecting extensive robot demonstrations.

State‑of‑the‑art results on mainstream simulation benchmarks

Conclusion: Xiaomi‑Robotics‑1 outperforms previous robotics foundation models across four widely used benchmarks.

Benchmark Xiaomi‑Robotics‑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 %

The gains are especially pronounced on RoboDojo, where the model’s ability to generalize to diverse manipulation scenarios shines.

Community reaction highlights excitement and practical concerns

Conclusion: The HN discussion reflects strong enthusiasm for household automation, while raising questions about hardware requirements and generalization.

"I'm not sure why some comments here are so pessimistic. I was grinning ear to ear while seeing the video. You're telling me we finally got robots that can do our laundry?" – user_7832

"Does this require a given hardware platform, or could I run this model with a cheap mini robot with a single camera?" – SillyUsername

"How well do models like these generalize to robot embodiments they weren’t trained on?" – mil22

These comments underscore two recurring themes:

  1. Desire for accessible hardware – users wonder whether the model can run on low‑cost platforms.
  2. Generalization across embodiments – the community is curious about transferability beyond Xiaomi’s proprietary robot.

Implications for the future of domestic robotics

Conclusion: By demonstrating that large‑scale, embodiment‑free pre‑training yields predictable performance gains, Xiaomi‑Robotics‑1 paves the way for a new generation of affordable, adaptable home robots.

If the scaling trend continues, future models could require even less task‑specific data, making personalized robot assistants feasible for a broad consumer base. The approach also suggests a roadmap for other robotics firms: invest in massive, automatically annotated video corpora, then perform lightweight embodiment alignment.

Citation

@article{guo2026xiaomi,
  title={Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories},
  author={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 Sun, Nan and others},
  journal={arXiv preprint arXiv:2607.15330},
  year={2026}
}

Sources