Xiaomi-Robotics-1 Breaks the Data Barrier in Robotics
The new robotics model pushes boundaries with large-scale data and training.
1The Data Barrier in Robotics
Access to large-scale, high-quality data in robotics has been a challenge, limiting the scalability of policy models. Xiaomi-Robotics-1 offers a new approach to overcome this barrier.
2Large-Scale Pre-Training
Xiaomi-Robotics-1 uses 100,000 hours of embodiment-free (UMI) pre-training data. This data covers over 1,700 scenarios and is processed with automated labeling.
3Real Robot Data
Post pre-training, the model is trained with real robot data. Xiaomi collected 7,200 hours of data in real homes to prepare the model for real-world tasks.
4Two-Stage Training
The model learns general representations during pre-training and then aligns with real robots and natural language commands. This makes the model versatile across a wide range of tasks.
5Adaptation to New Tasks
Xiaomi-Robotics-1 can efficiently learn new tasks. For example, it achieves a 75% success rate in tasks like phone packing or laundry loading.
6Simulation Success
The model achieves top results in four major simulation benchmarks, proving its generalization and scalability gains.