Mitigating the Human-Robot Domain Discrepancy in Visual Pre-training for Robotic Manipulation

lated tasks across two different benchmarks and five realworld tasks demonstrate significant improvements. These results span both single-task and language-conditioned multitask settings, evaluated using two different pre-trained models. Compared to existing pre-trained models, our adaptation method improves the average success rate by over 7% across multiple tasks on both simulated benchmarks and real-world evaluations. Project: https://jiamingzhou.github.io/projects/HumanRobotAlign