Brady Zhang
Brady Zhang is a researcher in the Department of Mechanical Engineering at University College London. He is best known as a co-author of VLA-AD, a distillation framework that compresses billion-parameter Vision-Language-Action policies into lightweight student models for real-time robotic control. His publication record also includes work on cycle-consistent generative adversarial networks for MR-only adaptive radiation therapy.
“Paper: 'Offline Semantic Guidance for Efficient Vision-Language-Action Policy Distillation' Authors: Jin Shi, Brady Zhang, Yishun Lu (UCL / Oxford)”
Source→“compressing a 7-billion-parameter VLA policy down to 158 million parameters while maintaining essentially identical task performance...matches the teacher with only a 0.27% average relative gap”
Source→“narrow error bars...consistent across the six teacher–suite rollouts, suggesting that it reflects a stable property of the manipulation data rather than a teacher-specific artifact”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.