AI suturing simulator trains surgical robots with 80% success rate

Surgical robotics has long struggled to simulate complete suturing scenarios, where soft tissue, rigid instruments, and deformable sutures interact simultaneously. Most existing simulators handle sutures or tissue in isolation, missing the critical contact dynamics during needle insertion. To address this, Tleukhan Mussin, Yafei Ou, and Mahdi Tavakoli (University of Alberta) developed a new simulation environment that models sutures using Position-Based Dynamics (PBD) and soft tissue using the Material Point Method (MPM), with a two-way contact coupling that accounts for frictional and drag forces. This allows for visually plausible suture-tissue interactions during insertion, driving, and extraction.

The simulator is optimized for GPU execution using multiple CUDA streams, enabling parallel scenes and fast training. The researchers integrated it as a Reinforcement Learning (RL) environment using Unity ML-Agents, where agents learned subtasks like needle insertion and extraction. Under the strictest distance threshold, RL agents achieved 80% success in needle insertion and 68% in extraction, demonstrating stable learning behavior. This work, accepted at BioRob 2026, represents a significant step toward high-fidelity, scalable simulation for autonomous surgical suturing, potentially accelerating the development of surgical robots that can perform delicate procedures with greater precision.