SutureAgent: Learning Surgical Trajectories via Goal-conditioned Offline RL in Pixel Space

A research team from institutions including Tsinghua University and MIT has developed SutureAgent, a novel AI framework that significantly improves the prediction of surgical needle trajectories from endoscopic video. The core innovation is formulating the task as a sequential decision-making problem, where the needle tip is treated as an agent moving step-by-step through pixel space. This approach naturally captures the continuity of motion and models physically plausible state transitions over time. To overcome the challenge of sparse expert annotations, the team uses cubic spline interpolation to generate dense reward signals, allowing the AI to exploit limited guidance while exploring plausible future paths.

SutureAgent employs a goal-conditioned offline reinforcement learning (RL) architecture. It encodes variable-length video clips to capture both spatial cues and long-term temporal dynamics, then autoregressively predicts future waypoints. For stable training from expert demonstrations alone, it uses Conservative Q-Learning with Behavioral Cloning regularization. The model was validated on a new, substantial dataset of kidney wound suturing containing 1,158 trajectories from 50 patients. The results are striking: SutureAgent achieved a 58.6% reduction in Average Displacement Error compared to the strongest existing baseline, demonstrating the effectiveness of its pixel-level sequential action learning paradigm for a critical task in robotic surgery.