Xiaomi Open Source VLA Large Model: Xiaomi-Robotics-0 Post-Training Process

Xiaomi's Robotics Division has open-sourced the entire training process for Xiaomi-Robotics-0 on a real device. Based on a pre-trained base, it can perform high-precision operations such as earphone storage using 20 hours of task data. The team overcame the model's "laziness effect" through a triple strategy of asynchronous inference, action prefixing and adaptive weighting, Λ-type masking, and random occlusion, achieving a balance between action continuity and visual sensitivity.