Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

STITCH: Augmented Dexterity for Suture Throws Including Thread Coordination and Handoffs

Kush Hari, Hansoul Kim, Will Panitch, Kishore Srinivas, Vincent Schorp, Karthik Dharmarajan, Shreya Ganti, Tara Sadjadpour, Ken Goldberg

TL;DR

STITCH tackles autonomous suturing by integrating a novel 6D needle pose estimation pipeline with an augmented dexterity motion controller that sequences needle insertion, thread sweeping, extraction, cinching, handover, and pose correction under closed-loop perception. The approach fuses stereo vision (RAFT-Stereo) and image segmentation (U-Net) with geometry-based pose estimation to enable reliable needle handling and thread management, achieving level 2 autonomy in a physical testbed. Experimental results show STITCH can autonomously perform an average of 2.93 sutures (4.47 with human supervision) over 15 trials, highlighting meaningful gains from perception–control integration and indicating pathways for robustness improvements. The work contributes a complete suturing pipeline, a 6D needle pose estimation method, and detailed physical-evaluation evidence that augmented dexterity can advance practical robotic suturing under surgeon supervision.

Abstract

We present STITCH: an augmented dexterity pipeline that performs Suture Throws Including Thread Coordination and Handoffs. STITCH iteratively performs needle insertion, thread sweeping, needle extraction, suture cinching, needle handover, and needle pose correction with failure recovery policies. We introduce a novel visual 6D needle pose estimation framework using a stereo camera pair and new suturing motion primitives. We compare STITCH to baselines, including a proprioception-only and a policy without visual servoing. In physical experiments across 15 trials, STITCH achieves an average of 2.93 sutures without human intervention and 4.47 sutures with human intervention. See https://sites.google.com/berkeley.edu/stitch for code and supplemental materials.

STITCH: Augmented Dexterity for Suture Throws Including Thread Coordination and Handoffs

TL;DR

STITCH tackles autonomous suturing by integrating a novel 6D needle pose estimation pipeline with an augmented dexterity motion controller that sequences needle insertion, thread sweeping, extraction, cinching, handover, and pose correction under closed-loop perception. The approach fuses stereo vision (RAFT-Stereo) and image segmentation (U-Net) with geometry-based pose estimation to enable reliable needle handling and thread management, achieving level 2 autonomy in a physical testbed. Experimental results show STITCH can autonomously perform an average of 2.93 sutures (4.47 with human supervision) over 15 trials, highlighting meaningful gains from perception–control integration and indicating pathways for robustness improvements. The work contributes a complete suturing pipeline, a 6D needle pose estimation method, and detailed physical-evaluation evidence that augmented dexterity can advance practical robotic suturing under surgeon supervision.

Abstract

We present STITCH: an augmented dexterity pipeline that performs Suture Throws Including Thread Coordination and Handoffs. STITCH iteratively performs needle insertion, thread sweeping, needle extraction, suture cinching, needle handover, and needle pose correction with failure recovery policies. We introduce a novel visual 6D needle pose estimation framework using a stereo camera pair and new suturing motion primitives. We compare STITCH to baselines, including a proprioception-only and a policy without visual servoing. In physical experiments across 15 trials, STITCH achieves an average of 2.93 sutures without human intervention and 4.47 sutures with human intervention. See https://sites.google.com/berkeley.edu/stitch for code and supplemental materials.
Paper Structure (25 sections, 4 figures, 1 table)

This paper contains 25 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Autonomous Suturing
  4. Visual Servoing in Surgical Automation
  5. Problem Statement
  6. Overview
  7. Assumptions
  8. Objectives and Evaluation Metrics
  9. Methods
  10. 6D Needle Pose Estimation Module
  11. Augmented Dexterity Suturing Motion Controller
  12. Needle Insertion
  13. Thread Sweeping
  14. Needle Extraction
  15. Suture Cinching
  16. ...and 10 more sections

Figures (4)

  • Figure 1: 6 sutures performed by STITCH. Step 1 shows Surgical Needle Insertion, Step 2 shows Needle Extraction, and Step 3 shows Needle Handover with Pose Correction. Detections of the needle endpoints are shown in the photo with the needle tip point shown as the orange circle and the needle swage point shown as the yellow circle. The light green circle represents the estimated needle pose.
  • Figure 2: 6D Needle Pose Estimation Module. The needle pose estimation starts with a pair of stereo left and right images. Using RAFT-Stereo, we generate a disparity image from the stereo pair lipson2021raft. Furthermore, we segment the needle in the left image with a U-Net to create a needle mask. From there, we apply the needle mask to the disparity image, and create the corresponding needle pointcloud. Using RANSAC, we find a best-fit plane to determine the normal vector of the 3D circle representing the needle (seen in green in the Fitted 3D Circle image). Then, we project all needle inliers from the RANSAC to the plane, and use RANSAC again to find the best fit circle (seen in black in the 3D Circle Fit image with the assumed fixed radius (12 mm for all experiments). Finally, we find the two farthest points on the needle pointcloud to determine the needle endpoints (seen as orange and yellow in the 6D Needle Pose image).
  • Figure 3: The individual processes embedded within the STITCH motion controller There are 4 parts to the state machine: 1. Needle insertion: (a) The right needle driver moves the needle to the initial insertion point at the proper orientation; (b) The right needle driver inserts the needle into the phantom with combined rotation and translation movements; 2. Sweeping and Needle Extraction with suture cinching: (c) The right needle driver follows the +y axis in the robot frame down the center of the wound to "sweep" any thread off the wound; (d) The left needle driver moves 1 centimeter behind the needle endpoint to prepare for extraction; (e) The left gripper grasps the needle and pulls it through until the length of the thread is at a desired $\beta$; 3. Needle Handover: (f) The right needle driver moves 1 centimeter behind the needle endpoint to prepare for handover; (g) The right gripper grasps the needle for handover; (h) The left gripper releases the needle; 4. Needle Pose Correction: (i) The right needle driver moves the needle to an optimal needle pose estimation region of the scene; (j) Based on the current pose of the needle, it is rotated such that the normal vector of the needle is aligned with the + y axis in robot frame; (k) The needle is rotated about the +y axis in robot frame so it is at the optimal orientation for the next insertion so the pipeline can be repeated again.
  • Figure 4: Histogram of the Number of Sutures to Failure by Method. The results in the histogram shown above is for 15 trials per each of the four methods.