Xiaomi-Robotics-1 shows that more data beats bigger models when training robots to move
Xiaomi released Xiaomi-Robotics-1, a foundational robot AI model that scales similarly to Large Language Models, with performance improving significantly as training data increases. To bypass the scarcity of robotic movement data, Xiaomi utilized handheld grippers with cameras to collect over 100,000 hours of motion recordings across diverse environments, avoiding expensive physical robot teleoperation. The dataset was efficiently labeled using an auxiliary AI model to generate text descriptions
Analysis
TL;DR
- Xiaomi released Xiaomi-Robotics-1, a foundational robot AI model that scales similarly to Large Language Models, with performance improving significantly as training data increases.
- To bypass the scarcity of robotic movement data, Xiaomi utilized handheld grippers with cameras to collect over 100,000 hours of motion recordings across diverse environments, avoiding expensive physical robot teleoperation.
- The dataset was efficiently labeled using an auxiliary AI model to generate text descriptions for each motion segment, allowing for scalable annotation in approximately two weeks.
- Empirical tests demonstrate that increasing training data yields greater performance gains than increasing model size, with success rates in unfamiliar environments rising from 25% to 75%.
- The model achieves state-of-the-art results on benchmarks like RoboCasa365 and RoboDojo, adapting to new tasks with less than ten hours of additional training data per task.
Why It Matters
This development addresses the critical data bottleneck in embodied AI, demonstrating that scalable, cost-effective data collection methods can rival traditional, expensive teleoperation techniques. It provides a clear roadmap for researchers and engineers to prioritize massive, diverse dataset accumulation over mere model parameter expansion for robotic generalization.
Technical Details
- Data Collection Strategy: Replaced physical robot arms with handheld grippers equipped with cameras, enabling data capture in real-world settings (kitchens, offices, outdoors) without dedicated robotic hardware.
- Automated Annotation: Utilized a separate AI model to automatically generate textual descriptions for over 100,000 hours of motion data, solving the labor-intensive labeling problem.
- Scaling Laws: Demonstrated that for vision-action models, data volume is a more significant driver of performance improvement than computational budget or model size, diverging from the balanced scaling of pure LLMs.
- Performance Metrics: Achieved a 75% average success rate on four distinct manipulation tasks (e.g., packing, laundry) with under ten hours of task-specific fine-tuning, outperforming competitors like Physical Intelligence’s pi0.7.
- Benchmark Leadership: Secured top positions on the RoboCasa365 and RoboDojo leaderboards, particularly excelling in unseen composite tasks and handling deformable materials.
Industry Insight
The industry must shift focus from purely architectural innovation to aggressive, scalable data acquisition strategies, as data diversity and quantity appear to be the primary limiting factors in current robotic AI capabilities. Companies should consider hybrid data collection methods, such as using consumer-grade hardware for broad environmental coverage, to build robust foundation models before deploying them on specialized physical platforms.
Disclaimer: The above content is generated by AI and is for reference only.