Path Planning for Flexible Needles Based on Multi-Agent Deep Reinforcement Learning

Flexible beveled tip needles offer distinct advantages for tumor ablation, but face challenges due to their nonlinear motion characteristics. Although single needle planning is effective for small tumors, larger tumors require multiple insertions, which significantly complicate the ablation process. Existing multi-needle path planning methods often assume static environments, which do not consider the dynamic movement of the deformable tissues. The proposed path planning scheme is simultaneously applicable to scenarios involving both needle-tissue interaction-coupled deformation and kinematic constraints of the multiple needles. By optimizing both the needle insertion point and trajectory, it achieves the optimality of the multiple flexible needle paths in dynamic compliant environments. Experimental results demonstrate that the proposed framework effectively simulates needle-tissue deformation, improving success rates by 35%-65% in multi-needle collaboration. Optimizing the needle entry point reduced the path length by more than 50% compared to random insertion points. This approach shows promise in enabling collaborative multi-needle insertion in robot-assisted minimally invasive surgery, where a multi-agent reinforcement learning mechanism coordinates needle motions and mitigates disturbances from dynamic tissue deformation.