Xiaomi Releases Xiaomi-Robotics-1: VLA Model Architecture Scaled on 100K Hours of Embodiment-Free Trajectories

Xiaomi-Robotics-1 overcomes physical robotics data bottlenecks by utilizing 100,000 hours of embodiment-free pre-training data aligned via a two-stage training pipeline. The resulting vision-language-action model demonstrates predictable performance scaling and state-of-the-art success rates in both physical deployment and simulation environments.

The Two-Stage VLA Training Pipeline

To overcome the physical data scarcity that historically caps robot policy scaling, Xiaomi-Robotics-1 utilizes a decoupled two-stage training paradigm: pre-training and post-training. This architecture separates broad representation learning from direct physical execution.

The first stage focuses on breadth by training the model on large-scale embodiment-free data, establishing general representations for action generation. The second stage, post-training, focuses on alignment. It maps these general action-generation capabilities onto physical robots through cross-embodiment datasets while converting the model's operational paradigm. Specifically, post-training shifts the model from generating actions based on raw scene state-transition descriptions to executing tasks directly from natural-language instructions.

Data Acquisition and Automated State-Transition Labeling

The pre-training phase relies on 100,000 hours of embodiment-free Universal Manipulation Interface (UMI) trajectories. This dataset spans over 1,700 scenarios across household, commercial, industrial, and outdoor environments. To process this volume of data without manual intervention, an automated annotation pipeline powered by a vision-language model (VLM) was engineered. The pipeline segments long video trajectories into fixed-length clips, and the VLM generates descriptions detailing the state transitions of grippers and interacting objects. This yields a massive text-action paired corpus.

Post-training transitions to high-quality cross-embodiment datasets. This includes 7,200 hours of in-house real-robot data collected in residential settings—specifically covering household manipulation tasks such as sofa tidying, shoe cabinet sorting, and kitchenware storage. This real-robot data is supplemented by filtered open-source robot data and manually annotated UMI trajectories. Unlike the auto-labeled pre-training data, this post-training UMI data is hand-labeled with precise temporal segments and instruction prompts.

Empirical Scaling and Physical Task Adaptation

Evaluation of the pre-training phase reveals a clean, predictable scaling law: validation action error steadily decreases as data volume and parameter size scale. Physical testing demonstrates that this scaling behavior transfers directly to real-world deployment. As pre-training parameters and data scale, the robot success rate in unseen environments with novel objects rises without showing signs of saturation.

For downstream tasks, the model exhibits high data efficiency during adaptation. Under restricted training budgets, Xiaomi-Robotics-1 significantly outperforms the established $\pi_{0.5}$ baseline.

• Phone Packing: 70% (vs. 30% baseline)
• Printer Refilling: 70% (vs. 20% baseline)
• Laundry Loading: 80% (vs. 40% baseline)
• Box Packing: 80% (vs. 70% baseline)

• Phone Packing: 80% (vs. 40% baseline)
• Printer Refilling: 60% (vs. 20% baseline)
• Laundry Loading: 100% (vs. 50% baseline)
• Box Packing: 100% (vs. 100% baseline)

Simulation Benchmark Performance

The generalization capability of Xiaomi-Robotics-1 was validated across four standard simulation suites. The model established new state-of-the-art results on all tested leaderboards.

On the RoboCasa benchmark, the model achieved a 74.5% success rate, representing a 2.6% relative gain over the second-best model's 72.6%. On RoboCasa365, it reached 57.4%, a 23.2% improvement over the second-best score of 46.6%. On VLABench, the model posted a 59.1% success rate, securing an 11.1% gain over the second-best score of 53.2%. On the RoboDojo suite, it recorded a 13.93% success rate, representing a 58.3% relative improvement over the second-place score of 8.80%.