Xuhui Kang

Pre-trained robot policies can be adapted to new tasks with drastically improved out-of-distribution generalization by anchoring fine-tuning to retrieved pre-training representations.