Xiaomi-Robotics-1 shows that more data beats bigger models when training robots to move

Xiaomi’s recent research on the Xiaomi-Robotics-1 system reveals that training with extensive datasets yields superior performance in robotic motion tasks compared to merely scaling up model sizes. The system was trained on over 100,000 hours of motion data, which was collected through human interactions using camera-equipped handheld grippers, rather than relying on robotic data collection methods. This approach emphasizes the importance of data quantity in enhancing the learning capabilities of robotic systems.

The findings indicate that the performance improvements from adding more data are substantial, suggesting that the model’s learning is heavily data-driven. Notably, the research highlights that the performance gains have not yet plateaued, indicating potential for further enhancements as more data becomes available. However, it is important to note that despite these improvements, the absolute success rates of the robotic movements remain low, pointing to ongoing challenges in the field of robotic motion training. For further details, refer to the original article on The Decoder.