EXIMO Improves Robot Policy Finetuning with VLM-Guided Exploration.

Current robotic manipulation policies rely heavily on behavior cloning from large VLA models, but finetuning these for new tasks remains difficult due to the high cost of data collection and the sample inefficiency of reinforcement learning (RL). EXIMO addresses this by proposing a three-stage approach. First, an exploration phase uses a VLM as a planner to break down complex tasks and guide data collection for the VLA. Second, an imitation phase finetunes the VLA using this newly orchestrated dataset. Finally, an optimization stage applies residual off-policy RL for further refinement. Experiments demonstrate that EXIMO significantly outperforms existing methods in both sample efficiency and final performance, making it a promising advancement for teaching robots new skills more quickly.