Xiaomi-Robotics-1: How to Work Around the Data Shortage for Robots

Language and vision models have long lived by empirical scaling laws: the more data, parameters and compute, the greater the capabilities. Robotics has been left out of the celebration — there is nowhere to get high-quality data in large volumes, and it is precisely this shortage, and nothing else, that has capped the growth of control models. Xiaomi proposes a workaround and describes it in the paper Xiaomi-Robotics-1.

Data Without a Body

The key idea is pretraining on “embodiment-free” trajectories in the UMI format. This means 100,000 hours of recordings covering more than 1,700 scenarios: homes, commercial spaces, industrial sites and open areas. Labeling such a corpus by hand is impossible, so Xiaomi built an automated pipeline: long videos are cut into fixed-length fragments, and a strong vision-language model describes how the state of the grippers and objects changes in each fragment.

During the fine-tuning stage, data from real robots comes into play — over 7,200 hours of recordings in actual apartments: cleaning a sofa, sorting out a shoe cabinet, laying out kitchen utensils. The scheme echoes the logic of training large language models: first pretraining for breadth, then alignment to specific robot “bodies” and to following instructions in natural language.

Scaling Works

The authors’ main conclusion is optimistic: the model scales cleanly. As the volume of data and the size of the model grow, the action-prediction error on the validation set steadily decreases, and after fine-tuning the share of successfully completed tasks on real robots rises too — even in unfamiliar settings and with unfamiliar objects. There are no signs of saturation yet, the team claims: the stronger the pretrained model, the better it works “out of the box.”

The Numbers

On new, complex tasks — packing phones, refilling a printer, loading a washing machine, stacking boxes — the model gets up to speed quickly. In less than 10 hours of demonstrations per task it reaches 75% success versus 40% for the baseline π0.5 at the same budget; with a budget of up to 40 hours the result climbs to 85%.

On four popular simulation benchmarks (RoboCasa, RoboCasa365, VLABench, RoboDojo), Xiaomi-Robotics-1 takes first place. Its lead over the nearest competitor ranges from a modest 2.6% on RoboCasa to an impressive 58.3% on RoboDojo.

In closing, the authors, following the spirit of such publications, promise an unedited recording of a robot packing a suitcase. The work has been published as a preprint on arXiv.

Source: Hacker News