100k‑Hour “Plug‑and‑Play” Robot Base Model: Xiaomi‑Robotics‑1 Tests Scaling Laws

Scaling Law Question for Robot Policy Models

Inspired by the predictable performance gains observed in large language and vision‑language models as data, model size, and compute increase, the authors ask whether robot policy models can exhibit similar scaling when trained on massive, diverse real‑world operation data.

Massive Real‑World Data Collection

During pre‑training, 100,000 hours of real‑world manipulation trajectories were captured using the Universal Manipulation Interface (UMI) across homes, commercial spaces, industrial settings, offices, and outdoors. The UMI system records detailed gripper‑state and object‑interaction changes without being tied to a specific robot body.

To handle this volume, an automated labeling pipeline splits each long trajectory into fixed‑length clips and employs a vision‑language model to generate textual descriptions of gripper and object state changes. This pipeline labeled the entire dataset in roughly two weeks.

Two‑Stage Training Paradigm

The model follows a pre‑training + post‑training approach.

Pre‑training: Learns a general action‑generation representation from UMI data. Given a visual observation and a language description, the model predicts an action sequence that transforms the scene toward the described goal.

Post‑training: Aligns the learned representation to real robot bodies (body alignment) and to natural‑language commands (instruction alignment). Approximately 10,000 hours of cross‑embodiment data were used, including 7,200+ hours from mobile and dual‑arm robots, 1,000+ hours of manually labeled UMI data, and public datasets such as Bridge V2, RT‑1, and DROID.

Scaling Experiments

Data‑scale experiments used 2.5K, 5K, 10K, and 20K hours of UMI data. Validation‑set action‑prediction loss consistently decreased as data grew, with smaller datasets showing over‑fitting and larger datasets yielding more stable training.

Model‑size experiments compared 2 B, 5 B, and 10 B‑parameter versions. Larger models continuously improved prediction accuracy, confirming that both data and model scale benefit performance.

Out‑of‑the‑Box Capability

After post‑training, the model can execute a variety of real‑world tasks (e.g., sofa tidying, utensil storage, shoe‑cabinet organization) directly from natural‑language commands, without task‑specific fine‑tuning.

In real‑robot evaluations, Xiaomi‑Robotics‑1 outperformed the Pi‑0.5 baseline across four tasks, even when each task had less than 10 hours of downstream data, demonstrating efficient new‑task adaptation.

State‑of‑the‑Art Simulation Benchmarks

RoboCasa365: 57.4% average success (previous best 46.6%). Strong performance on the Composite‑Unseen split.

RoboDojo: 20.07 average score and 13.93% success, surpassing the prior best of 13.07/8.80%.

VLABench: 59.1% success, 70.3% progress score.

RoboCasa: 74.5% success, beating methods such as RLDX‑1, Cosmos Policy, GR00T N1.6, Pi‑0.5, and Pi‑0‑FAST.

Scalable Training Path for Embodied Intelligence

Large‑scale pre‑training to learn universal action representations.

Cross‑embodiment post‑training to transfer capabilities to real robots.

Few‑shot downstream fine‑tuning for rapid adaptation to complex new tasks.

This pathway suggests robot policy models can move from small, task‑specific datasets toward a foundation‑model paradigm, with scaling laws still in early exploration but showing promising performance gains as data, model size, and task coverage expand.

Code and model weights will be released gradually.