Xiaomi Open-Sources Full Post-Training Process of VLA Large Model Xiaomi-Robotics-0
Published · Apr 27 · Mon Source · 小米技术 (CN)

Xiaomi Open-Sources Full Post-Training Process of VLA Large Model Xiaomi-Robotics-0

Xiaomi Robotics Division open-sources the full post-training process for Xiaomi-Robotics-0 on real machines. Based on a pre-trained base, it can complete high-precision operations such as earphone storage using 20 hours of task data. The team addressed the model's "lazy effect" through three strategies: asynchronous inference, action prefix and adaptive weighting, and Λ-shaped masking and random masking, achieving a balance between action coherence and visual sensitivity.

KeywordsXiaomiOpen-SourcesFullPost-TrainingProcessVLALargeModel

Xiaomi Robotics Division open-sources the full post-training process for Xiaomi-Robotics-0 on real machines. Based on a pre-trained base, it can complete high-precision operations such as earphone storage using 20 hours of task data. The team addressed the model's "lazy effect" through three strategies: asynchronous inference, action prefix and adaptive weighting, and Λ-shaped masking and random masking, achieving a balance between action coherence and visual sensitivity.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.