Xiaomi Robotics 1 — a robot foundation model trained on 100,000 hours of embodiment-free trajectories, showing clean scaling laws, VLM-powered auto-labeling, and efficient adaptation to new tasks with under 10 hours of demos reaching 75% success rate.

Xiaomi has unveiled Robotics 1 — its first foundation model for humanoid robots, breaking through the main bottleneck in modern robotics: data scarcity. Rather than relying solely on expensive, slow real-robot data collection, Xiaomi adopted a two-stage training approach that mirrors the paradigm shift seen in large language models.

100,000 Hours of Embodiment-Free Pre-Training

In the pre-training stage, Xiaomi Robotics 1 consumes 100,000 hours of UMI (Unified Manipulation Interface) trajectories spanning over 1,700 scenarios — household, commercial, industrial, and outdoor spaces. This gives the model broad understanding of the physical world without needing to control a real robot.

Xiaomi built a VLM-powered auto-labeling pipeline that splits videos into fixed-length clips and describes state transitions of grippers and interacting objects. This solves the scaling bottleneck of manual annotation — one of the biggest obstacles in robotics.

The researchers found that pre-training shows clean scaling behavior: as data and model size grow, validation action error steadily decreases. More data directly translates to better outcomes.

7,200 Hours of Real-Robot Post-Training

The post-training stage aligns the model along two axes. Embodiment alignment uses high-quality cross-embodiment real-robot data to map the general action-generation ability onto actual robots. Instruction alignment shifts the model from acting on state-transition descriptions to understanding and executing natural-language instructions.

The results are striking: real-robot success rate rises predictably as pre-training data and model size increase. No signs of saturation have been observed — improvement continues with scaling.

Efficient Adaptation — Under 10 Hours to Learn

Xiaomi Robotics 1 can learn new tasks with exceptional data efficiency. With an average of under 10 hours of demonstrations per task, it already reaches 75% overall success rate — nearly double that of competing models like π0.5. This makes it practical for real-world business deployment.

Benchmark Results

The model achieves State-of-the-Art results on CALVIN, LIBERO, and SIMPLER simulation benchmarks. On CALVIN, it scores 88.3%, outperforming all previous models and validating the two-stage approach.

What This Means for Georgia

Xiaomi's open approach to robotics means developing countries, including Georgia, can access cutting-edge robot technology. Logistics automation, agricultural robots, and service industry robotics are areas where Georgia can leverage this technology for economic growth.

Conclusion

Xiaomi Robotics 1 is more than just a new robot model. It is proof that scaling laws work in robotics the same way they do in LLMs. 100,000 hours of embodiment-free pre-training, VLM auto-annotation, and two-stage training create a new standard that other companies will soon follow.

📖 Source