a day ago
- Xiaomi-Robotics-1 breaks the robotics data bottleneck by scaling policy models with embodiment-free pre-training, using 100,000 hours of UMI data.
- The model uses a two-stage training paradigm: pre-training learns general action generation from UMI data, while post-training aligns with real robots and instruction following.
- Scaling laws hold: increasing pre-training data and model size reduces validation error and boosts real-robot success rates after post-training without saturation.
- The model achieves high data efficiency for new tasks, reaching 75% success with under 10 hours of demonstrations per task, and sets state-of-the-art results on four simulation benchmarks.
- Applications include efficient adaptation to complex real-robot tasks and superior performance in simulation benchmarks like RoboCasa and VLABench.