Xiaomi's Robotics Model Xiaomi-Robotics-1 Breaks the Data Barrier
A groundbreaking model that enhances robotic capabilities through large-scale data training.
1The Data Barrier in Robotics
The lack of large-scale, high-quality data in robotics has limited the scalability of policy models. Xiaomi-Robotics-1 overcomes this barrier by combining extensive pre-training with real-robot data. This approach pushes the boundaries of robotic capabilities.
2Data Collection and Processing
Xiaomi-Robotics-1 utilizes 100,000 hours of pre-training data and 7,200 hours of real-robot data. The data is collected from household, commercial, and industrial environments, covering diverse tasks. An automated labeling pipeline enriches the data with language-based descriptions.
3Training Process
The model undergoes a two-stage training process: pre-training and post-training. Pre-training equips the model with broad action-generation capabilities, while post-training aligns it with real robots. This process enables the model to understand natural language commands.
4Scalability and Performance
Xiaomi-Robotics-1 demonstrates consistent performance improvement as pre-training data and model size increase. The real-robot success rate mirrors the scaling behavior observed in pre-training, indicating better performance with more data.
5Application Areas
The model can learn new tasks with just a few hours of data. It achieves a 75% success rate in complex tasks like phone packing and laundry loading. In simulation benchmarks, it outperforms competitors, setting new standards in robotic performance.
6A New Era in Robotics
Xiaomi-Robotics-1 showcases the potential of large-scale data and scalable models in robotics. This advancement could revolutionize household and industrial robotic applications. Xiaomi's innovation paves the way for future robotic technologies.