Scheduling Adaptive Imitation Learning for Long-Horizon Dexterous Robot Micromanipulation of Deformable Cell

Robots performing collaborative, long-horizon dexterity cell micromanipulation tasks are challenging and practically significant, such as stripping intact cell membranes, which is considered as one of the most technically demanding procedures. Imitation learning approach is expected to address the challenges of multi-task coupling and object modeling difficulties in long-horizon tasks. Existing imitation learning algorithms have significant shortcomings in long-horizon tasks due to the presence of compound errors, as performing only a simple mapping of the task environment space to the action space without adaptive action scheduling. In this paper, we propose scheduling adaptive imitation learning (SAIL) method for long-horizon dexterous robot micromanipulation tasks. The algorithm can adaptively output sparse action sequences and single-step interactive actions according to the complexity of the task. Physical experiments show that the SAIL algorithm can complete the deformable micrometer-scale zebrafish embryonic cells dexterous membrane stripping surgery. Ablation studies further confirmed SAIL's high efficiency in various subtasks including PushCell, PeelCell, ExtrudeCell, and Rotate, ultimately achieving an average accuracy of 84.4%, significantly outperforming existing methods.