SutureAgent: Learning Surgical Trajectories via Goal-conditioned Offline RL in Pixel Space
Summary: arXiv:2603.26720v1 Announce Type: cross
Abstract
Predicting surgical needle trajectories from endoscopic video is critical for robot-assisted suturing, enabling anticipatory planning, real-time guidance, and safer motion execution. Existing methods that directly learn motion distributions from visual observations tend to overlook the sequential dependency among adjacent motion steps. Moreover, sparse waypoint annotations often fail to provide sufficient supervision, further increasing the difficulty of supervised or imitation learning methods.
Introduction
In the realm of robotic surgery, accurate prediction of needle trajectories is essential for enhancing the performance and safety of surgical procedures. Traditional approaches often face significant challenges due to the inherent complexity and variability in surgical environments. To tackle these issues, we introduce a novel methodology that reformulates the trajectory prediction problem.
Methodology
We propose a framework named SutureAgent, which treats the needle tip as an agent that moves step by step in pixel space. This unique formulation captures the continuity of needle motion and allows for the explicit modeling of physically plausible pixel-wise state transitions over time. The following aspects are key to our approach:
- Sequential Decision-Making: The trajectory prediction is framed as a sequential decision-making problem, which better represents the nature of surgical actions.
- Goal-Conditioned Reinforcement Learning: By employing a goal-conditioned offline reinforcement learning approach, SutureAgent is able to leverage sparse annotations effectively.
- Cubic Spline Interpolation: This technique transforms sparse waypoint annotations into dense reward signals, thereby facilitating a more robust learning process.
- Observation Encoder: Variable-length clips are encoded to capture both local spatial cues and long-range temporal dynamics.
- Action Prediction: Future waypoints are predicted autoregressively through actions that consist of discrete directions and continuous magnitudes.
Implementation
To ensure stable offline policy optimization from expert demonstrations, we adopt Conservative Q-Learning combined with Behavioral Cloning regularization. This hybrid approach enables SutureAgent to effectively learn from limited expert data while mitigating overfitting risks.
Results
We conducted experiments on a new kidney wound suturing dataset containing 1,158 trajectories from 50 patients. The results demonstrated a significant improvement in performance, with SutureAgent achieving a 58.6% reduction in Average Displacement Error compared to the strongest baseline. This remarkable achievement underscores the efficacy of modeling needle trajectory prediction as pixel-level sequential action learning.
Conclusion
The introduction of SutureAgent marks a pivotal advancement in the field of robotic surgery. By addressing the limitations of previous methods and effectively utilizing sparse annotations, this framework paves the way for more accurate and reliable surgical procedures. Future work will involve further refinement of the model and exploration of its applicability across various surgical tasks.