SutureBot: A Precision Framework & Benchmark For Autonomous End-to-End Suturing

TL;DR

The paper tackles autonomous end-to-end suturing, a long-horizon dexterous manipulation problem, by introducing SutureBot—a precision-focused benchmark on the dVRK with a 1,890-demo dataset and a goal-conditioned imitation-learning framework. It evaluates multiple vision-language-action models and a high-level language policy, showing that goal conditioning significantly improves insertion-point precision (approximately $1.0-1.3$ mm) and overall targeting accuracy, while end-to-end success remains challenging (ACT achieves 3/10 end-to-end trials). The contributions provide a reproducible benchmark, a large real-world suturing dataset, and insights into how goal representations and pretraining affect performance in dexterous, long-horizon robotic surgery. Together, these results pave the way for more robust, precise autonomous suturing and set a foundation for future data- and model-driven advances in robotic-assisted procedures.

Abstract

Robotic suturing is a prototypical long-horizon dexterous manipulation task, requiring coordinated needle grasping, precise tissue penetration, and secure knot tying. Despite numerous efforts toward end-to-end autonomy, a fully autonomous suturing pipeline has yet to be demonstrated on physical hardware. We introduce SutureBot: an autonomous suturing benchmark on the da Vinci Research Kit (dVRK), spanning needle pickup, tissue insertion, and knot tying. To ensure repeatability, we release a high-fidelity dataset comprising 1,890 suturing demonstrations. Furthermore, we propose a goal-conditioned framework that explicitly optimizes insertion-point precision, improving targeting accuracy by 59\%-74\% over a task-only baseline. To establish this task as a benchmark for dexterous imitation learning, we evaluate state-of-the-art vision-language-action (VLA) models, including $π_0$, GR00T N1, OpenVLA-OFT, and multitask ACT, each augmented with a high-level task-prediction policy. Autonomous suturing is a key milestone toward achieving robotic autonomy in surgery. These contributions support reproducible evaluation and development of precision-focused, long-horizon dexterous manipulation policies necessary for end-to-end suturing. Dataset is available at: https://huggingface.co/datasets/jchen396/suturebot

Figures (6)

• Figure 1: Overview of the precision-conditioned control framework for long-horizon, dexterous surgical tasks. Image observations are processed by a high-level language policy, which selects the current task and generates the associated language condition. The user specifies target needle insertion and exit points via a graphical interface, which is used to generate the goal condition. These inputs, language condition, goal condition, and real-time kinematic data, are then processed by the low-level policy to produce precise, continuous control commands for the robot.
• Figure 2: Experimental setup showing the Da Vinci Research Kit (dVRK), remote center of motion fixture, and suture pad. The robot, suture pad, and task utilized for data collection was selected to allow others to reproduce the benchmark. All data is collected on wound one of the suture pad, while wounds two through six are used for generalization testing.
• Figure 3: The suturing procedure is broken into three tasks, needle pickup, needle throw, and knot tie. These task discretizations were then utilized for data collection, policy training, and evaluation.
• Figure 4: To enable a policy to approach targeted points, we utilize goal conditions generated from target points, which serve as inputs to the model during training and inference. We evaluate three types of goal conditions; point labels on the endoscope image, an additional image input with masks, and an additional image input with a distance map.
• Figure 5: Ultraviolet (UV) marks were utilized to measure the accuracy of the policy. After marking the wound, the target points were selected in the endoscope view using a UV light. During execution, the UV light was off and the marks were invisible. After the task, the UV light was turned on and the distance from the suture to the mark was measured.