SutureBot: Robotic Suturing Systems

• SutureBot is a suite of robotic suturing systems that integrate automated needle pickup, localization, tissue penetration, and knot tying under strict geometric constraints.
• It employs varied actuation methods including tetherless magnetic devices and traditional articulated robots to achieve precise, controlled suturing tasks.
• Research leverages advanced imaging, vision-based state estimation, and closed-loop control to enhance suturing precision and benchmark system performance.

SutureBot is a term used in recent surgical-robotics literature for several related but non-identical systems, pipelines, and benchmarks concerned with robotic suturing. Across these usages, the central problem is the same: coordinated automation or assistance for needle pickup, needle localization, tissue penetration, suture-thread handling, pull-through, and knot tying under stringent geometric and perceptual constraints. The term appears in work on impact-based magnetic tetherless suturing, neonatal and pediatric teleoperation, OCT-guided vascular anastomosis, image-to-grasp thread manipulation, and a dVRK end-to-end benchmark, rather than denoting a single universally standardized platform (Erin et al., 2021, Marinho et al., 2021, Joglekar et al., 2024, Haworth et al., 2024, Haworth et al., 23 Oct 2025).

1. Terminological scope and historical development

The research lineage associated with SutureBot begins with subtask-specific automation rather than complete suturing systems. Early work addressed suture-thread detection via a multi-stage deep-learning pipeline that produced a gradient road map and overlap map, then linked curvilinear segments into a full thread representation (Hu et al., 2017). Needle pickup was studied as a separate vision-guided capability, including a da Vinci-based system that tracked three green fiducial markers on a 1/2-circle needle and autonomously executed an approach and grasp sequence (D'Ettorre et al., 2018). OCT-guided micro-suturing then introduced automatic calibration, wound detection, keypoint identification, and path planning for sub-millimeter needle guidance through soft tissue (Tian et al., 2020).

Later papers use “SutureBot” more explicitly to describe broader robotic suturing architectures. An impact-force-based magnetic suturing mechanism was presented as an “Impact‐Force‐Based Magnetic Suturing System for SutureBot Design” (Erin et al., 2021). The SmartArm neonatal suturing feasibility study described the configured platform as a neonatal suturing robot (“SutureBot”) (Marinho et al., 2021). A 2024 image-to-grasp thread-manipulation pipeline was framed as the “SutureBot pipeline” (Joglekar et al., 2024). The Micro Smart Tissue Autonomous Robot, or µSTAR, was described as “often referred to as SutureBot” (Haworth et al., 2024). In 2025, “SutureBot” became the title of a precision framework and benchmark for autonomous end-to-end suturing on the dVRK (Haworth et al., 23 Oct 2025). This suggests an evolution from isolated perception and manipulation primitives toward integrated long-horizon formulations.

2. Robotic embodiments and actuation strategies

The systems labeled or associated with SutureBot span substantially different actuation regimes. One branch uses tetherless magnetic actuation: a 3-DoF planar needle mechanism in which planar positioning $(X,Y)$ and in-plane rotation $(\theta)$ are generated by a four-coil electromagnetic array, while an internal traveling permanent magnet acts as a “magnetic hammer” to deliver momentary high-impact forces for tissue penetration (Erin et al., 2021). Another branch uses conventional articulated surgical robots such as the dVRK, KUKA LBR Med, ABB IRB-120, and DENSO VS050, often augmented with endoscopes, wrist cameras, OCT probes, and specialized suturing tools (Tian et al., 2020, Marinho et al., 2021, Haworth et al., 2024, Haworth et al., 23 Oct 2025).

Platform	Hardware basis	Distinctive feature
Magnetic SutureBot design	Four-coil electromagnetic array; 12 G stainless-steel cannula	Internal NdFeB “piston” magnet for impact penetration (Erin et al., 2021)
SmartArm SutureBot	Two DENSO VS050 manipulators; 3.5 mm instruments	QP-based teleoperation in neonatal workspace (Marinho et al., 2021)
µSTAR / SutureBot	KUKA LBR Med; modified Endo360; OCT sensor; microcamera	Autonomous vascular anastomosis on small-diameter vessels (Haworth et al., 2024)
End-to-end SutureBot benchmark	dVRK Si; stereo endoscope; dual wrist cameras	1,890-demonstration benchmark with goal-conditioned policies (Haworth et al., 23 Oct 2025)

The magnetic system is distinctive because it explicitly addresses the force-scaling problem of miniature magnetic end effectors. Its optimized custom-built 12 G needle generated $1.16\ \mathrm{N}$ penetration force, reported as 56 times stronger than magnetic counterparts of the same size without impact force, while keeping the overall needle motion slow and easily controllable (Erin et al., 2021). The detailed design used a standard 12 G stainless-steel cannula with outer diameter $2.77\ \mathrm{mm}$, a cylindrical NdFeB magnet of diameter $1.59\ \mathrm{mm}$ and length $12.70\ \mathrm{mm}$, a PTFE sleeve with friction of approximately $10\text{–}20\ \mathrm{mN}$, and a total assembly mass of about $0.15\ \mathrm{g}$ (Erin et al., 2021). Peak impact force was modeled as

$$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$

with an optimal balance obtained by setting $l_m=2l_{\rm tube}/3$ under the stated constraint (Erin et al., 2021).

At the opposite end of the design spectrum, µSTAR integrates a modified Endo360 laparoscopic suturing instrument, a brushless DC motor, a 100 kHz swept-source OCT engine at center wavelength $(\theta)$0, axial resolution $(\theta)$1 in tissue, and an Omnivision OV6946 microcamera with $(\theta)$2 px output at up to 60 fps (Haworth et al., 2024). The 2025 dVRK benchmark platform combines a stereo endoscope at $(\theta)$3 px and 30 Hz with two $(\theta)$4 px wrist cameras mounted $(\theta)$5 from each wrist, and logs images and 6-DoF kinematics at 30 Hz (Haworth et al., 23 Oct 2025). SmartArm, by contrast, emphasizes access in a neonatal cavity through two 20 cm-long, 3.5 mm-diameter instruments providing 9 DoF per arm and a passive 4 mm rigid endoscope (Marinho et al., 2021).

3. Needle localization and grasp-state estimation

Reliable needle-state estimation is a foundational SutureBot capability because the needle is thin, reflective, frequently occluded, and often kinematically coupled to a gripper. A central result in this literature is that independent reconstruction of needle and gripper can be physically inconsistent. The constrained 6D in-hand needle tracking framework therefore replaces a conventional 6D pose state $(\theta)$6 with a 4D reparameterization

$(\theta)$7

where $(\theta)$8 indexes the grasp point on the circular needle, $(\theta)$9, $1.16\ \mathrm{N}$0, and $1.16\ \mathrm{N}$1, so that the feasible grasp set becomes a hyperrectangle $1.16\ \mathrm{N}$2 in the new state space (Chiu et al., 2022). This allows feasible contact constraints to be enforced directly inside Bayesian filters rather than by post hoc rejection of infeasible 6D hypotheses.

In simulation, the reparameterized particle filter $1.16\ \mathrm{N}$3 achieved approximately $1.16\ \mathrm{N}$4 and $1.16\ \mathrm{N}$5 median errors even at $1.16\ \mathrm{N}$6 px image noise, while the unconstrained particle filter drifted to approximately $1.16\ \mathrm{N}$7 and $1.16\ \mathrm{N}$8 (Chiu et al., 2022). The reparameterized filter also ran at approximately $1.16\ \mathrm{N}$9, whereas the rejection-sampling constrained particle filter ran at approximately $2.77\ \mathrm{mm}$0 (Chiu et al., 2022). In ex vivo stereo-endoscope experiments on dVRK hardware, the constrained method always produced a physically valid grip, while the unconstrained filter could yield needle reconstructions that “float” or collide with the jaws (Chiu et al., 2022).

Complementary needle-localization work has focused on autonomous regrasp and handover. HOUSTON addressed handoff of unmodified, surgical, tool-obstructed needles using a learned active sensing policy, stereo segmentation, triangulation, RANSAC circle fitting, and multi-camera high-precision grasping (Wilcox et al., 2021). On the dVRK, it achieved 96.3% success, corresponding to $2.77\ \mathrm{mm}$1 successful left-to-right and right-to-left handovers over 28 grasp configurations, with 21–24 s per handover (Wilcox et al., 2021). In multi-handover experiments with $2.77\ \mathrm{mm}$2, the “towards-camera” configuration averaged 38.6 handovers before failure and the “away-camera” configuration averaged 26.2 (Wilcox et al., 2021). On unseen needles, reported success ranged from 75% to 92.9% (Wilcox et al., 2021).

Magnetic suturing introduces a different localization regime because control is mediated by nonlinear fields and the scene can be cluttered by blood and tissue. In that setting, a custom U-Net with a MobileNetV2-style encoder, augmented by DBSCAN clustering, RANSAC line fitting, and visibility classification heads, achieved $2.77\ \mathrm{mm}$3 RMS error in clean environments, $2.77\ \mathrm{mm}$4 RMS error in challenging environments with blood and occlusion, and $2.77\ \mathrm{mm}$5 average localization RMS over all environments (Pryor et al., 2021). Combined with closed-loop control, the magnetic needle followed a running-suture path with tip-position tracking errors ranging from $2.77\ \mathrm{mm}$6 to $2.77\ \mathrm{mm}$7 RMS across four environmental conditions (Pryor et al., 2021). Earlier marker-based pickup work reached $2.77\ \mathrm{mm}$8 overall autonomous approach-and-grasp success, but required three manually affixed green fiducial markers and user-defined ROIs in the first frame (D'Ettorre et al., 2018).

4. Suture-thread reconstruction and image-to-grasp pipelines

Thread perception is a second major SutureBot subfield, distinct from needle pose estimation because the thread is thin, deformable, partially observed, and topologically variable. One line of work uses transfer learning and graph search. A learning-driven framework with spatial optimization fine-tuned a U-Net using 200 stereo image pairs with synthetic threads for offline pretraining and 40 on-site images for online adaptation, then performed pixel-level ordering and 3D reconstruction by an optimized shortest-path method (Lu et al., 2020). On dVRK test sets, the transfer-learning model $2.77\ \mathrm{mm}$9 achieved 71.1% IoU, 72.5% precision, and 97.3% recall on Set 1, and 72.1% IoU, 73.5% precision, and 97.5% recall on Set 2 (Lu et al., 2020). The complete stereo-frame pipeline ran in less than 70 ms, and 100% of runs achieved valid grasp points within the 10 mm robot grasper aperture (Lu et al., 2020).

A second line emphasizes reliability-aware geometry. Reliability-driven keypoint detection and Minimum Variation Spline smoothing reconstructed a 3D centerline from stereo images by retaining only reliable keypoints, fitting a degree-4 B-spline, and constraining the spline with depth envelopes derived from local fit deviations (Joglekar et al., 2022). On 40 simulated scenes, the reported mean 3D curve error was $1.59\ \mathrm{mm}$0, the maximum error was $1.59\ \mathrm{mm}$1, the length error was $1.59\ \mathrm{mm}$2, and there were $1.59\ \mathrm{mm}$3 failures, three of them optimizer-related (Joglekar et al., 2022). On real surgical pig data, mean reprojection error was approximately 0.5 px and subjective quality was at least $1.59\ \mathrm{mm}$4 in 8 of 10 cases (Joglekar et al., 2022).

A third line uses temporal tracking and self-supervision. For autonomous suture tail-shortening, self-supervised learning under UV illumination trained a U-Net detector on 1,320 images and reconstructed the thread as a NURBS spline from stereo triangulation and subsequent tracking updates (Schorp et al., 2023). The method reported 1.33 px average reprojection error on single-frame 3D reconstructions, 0.84 px average reprojection error on two tracking sequences, and 90% success across 20 tail-shortening trials with mean absolute tail error $1.59\ \mathrm{mm}$5 and mean time approximately 107 s (Schorp et al., 2023).

The most direct image-to-grasp SutureBot formulation is the 2024 reliability-driven thread-reconstruction and grasping pipeline. It segments the thread with HQ-SAM, rejects outliers using a stereo-match-cost criterion, defines per-observation 3D reliability regions $1.59\ \mathrm{mm}$6, and fits a cubic B-spline $1.59\ \mathrm{mm}$7 by solving an iterative linearly constrained quadratic program under the minimum-variation objective $1.59\ \mathrm{mm}$8 (Joglekar et al., 2024). Grasping is then posed as a probabilistic “capture, slide, grasp” policy with

$1.59\ \mathrm{mm}$9

and the capture point is selected by maximizing $12.70\ \mathrm{mm}$0 over sampled candidates (Joglekar et al., 2024). Over 400+ total trials, direct grasping achieved 90.5% overall success, while robust CSG grasping achieved 97.0% success, with average robust-grasp execution time of approximately 8 s and a 6.5% absolute gain over direct grasping (Joglekar et al., 2024). The paper states that this gain confirms recovery from spline errors exceeding 5–10 mm (Joglekar et al., 2024).

5. Control architectures and execution regimes

SutureBot control spans teleoperation with active constraints, shared autonomy, conditional autonomy, and autonomous subprocedures. In confined pediatric or neonatal workspaces, quadratic-programming control is a recurring design pattern. SmartArm solves, at each 1 kHz control step, a QP in joint velocity space with translation and rotation tracking terms, motion penalties, and linear inequalities that encode joint-limit avoidance and an entry-sphere constraint (Marinho et al., 2021). With $12.70\ \mathrm{mm}$1, $12.70\ \mathrm{mm}$2, $12.70\ \mathrm{mm}$3, and Phantom Premium haptic devices on the operator side, a medically inexperienced operator completed all ten two-throw intracorporeal knots in a neonatal chest model (Marinho et al., 2021). Task times ranged from 216 s to 1421 s, with mean approximately 502 s and median approximately 437 s (Marinho et al., 2021).

Virtual-fixture assistance extends the same constrained-optimization logic from basic teleoperation to guided suturing maneuvers. Looping virtual fixtures and a trajectory-guidance cylinder were implemented through QP objectives and Cartesian force feedback so that the second tool remains inside a dynamic tubular region around the first instrument’s shaft (Marinho et al., 2019). In simulation with six engineering students, median completion times were 14.56 s for entry-sphere only, 17.12 s for entry plus shaft–shaft avoidance, and 17.10 s for entry plus shaft–shaft avoidance plus LVFs plus TGC, while mean loop-surface error improved from 2.27 mm to 1.21 mm and only the baseline condition collided (Marinho et al., 2019). In physical experiments with pediatric surgeons, mean loop error improved from 0.89 mm to 0.59 mm for the expert user and from 2.16 mm to 1.92 mm for the intermediate user (Marinho et al., 2019).

OCT-guided autonomous suturing adds tissue-specific geometric planning. In micro-suturing with an ABB IRB-120, calibration jointly estimated $12.70\ \mathrm{mm}$4 and $12.70\ \mathrm{mm}$5 from $12.70\ \mathrm{mm}$6 needle poses by minimizing a Frobenius-norm least-squares objective, while wound analysis extracted start, end, and deepest points from OCT B-scans and planned a circular needle arc with desired suture depth $12.70\ \mathrm{mm}$7 and $12.70\ \mathrm{mm}$8 (Tian et al., 2020). On tissue phantoms and porcine tissue, the system reported 3D RMSE of 0.177 mm and 0.219 mm respectively, with overall average 3D RMSE approximately 0.20 mm, wound segmentation and keypoint-detection success of 98.6% over 590 B-scans, and a final demonstration of three sequential throws on porcine skin (Tian et al., 2020).

Conditional autonomy in soft-tissue suturing was demonstrated by STAR for intestinal anastomosis. STAR operates at level of autonomy 3 out of 5, generates multiple suture plans, waits for tissue to become stationary, and forces replanning if the deformation score $12.70\ \mathrm{mm}$9 exceeds $10\text{–}20\ \mathrm{mN}$0 (Saeidi et al., 2021). Ex vivo, STAR reported $10\text{–}20\ \mathrm{mN}$1, spacing $10\text{–}20\ \mathrm{mN}$2 with $10\text{–}20\ \mathrm{mN}$3, bite $10\text{–}20\ \mathrm{mN}$4 with $10\text{–}20\ \mathrm{mN}$5, and $10\text{–}20\ \mathrm{mN}$6 (Saeidi et al., 2021). In vivo porcine experiments reported $10\text{–}20\ \mathrm{mN}$7, $10\text{–}20\ \mathrm{mN}$8, $10\text{–}20\ \mathrm{mN}$9, and $0.15\ \mathrm{g}$0 (Saeidi et al., 2021).

The magnetic impact-force system represents a different execution regime: teleoperated but mechanically specialized for penetration. Using a square-wave pulling-and-pushing sequence with period $0.15\ \mathrm{g}$1, duty ratio $0.15\ \mathrm{g}$2, and typical coefficients $0.15\ \mathrm{g}$3, $0.15\ \mathrm{g}$4, it achieved peak forces exceeding 1 N for $0.15\ \mathrm{g}$5, piston peak velocity $0.15\ \mathrm{g}$6, impact rise time of approximately 5–10 ms, and repeatability of $0.15\ \mathrm{g}$7 over 5 runs (Erin et al., 2021). It penetrated 1.6 mm bacon to a depth of 11.8 mm in 16 s over 20 hammer cycles, and completed a teleoperated running stitch with three penetrations in 158 s on a 0.6% agarose phantom covered by surgical gauze (Erin et al., 2021).

6. Benchmarks, comparative performance, and unresolved challenges

The most explicit attempt to formalize SutureBot as an end-to-end benchmark is the 2025 dVRK framework built around needle pickup, tissue insertion, and knot tying. It provides 1,890 teleoperated demonstrations, including 628 pickups, 310 throws, 952 knots, and 454 recovery episodes, recorded with stereo endoscope images, dual wrist-camera streams, and 30 Hz kinematics (Haworth et al., 23 Oct 2025). The architecture is hierarchical: a high-level policy predicts language conditions $0.15\ \mathrm{g}$8, while a low-level vision-language-action policy receives images, recent kinematics, and user-specified pixel goals $0.15\ \mathrm{g}$9 and outputs $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$0 (Haworth et al., 23 Oct 2025).

Under this protocol, ACT reported $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$1 pickup success, $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$2 throw success, $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$3 pull-through success, $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$4 knot-tie success, insertion error $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$5, exit error $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$6, time $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$7, and $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$8 end-to-end success (Haworth et al., 23 Oct 2025). $$F_{\rm impact}^{\max} \approx \sqrt{\frac{2\,m\,F_m\,d}{(\Delta t_{\rm impact})^2}},$$9 reached $l_m=2l_{\rm tube}/3$0, $l_m=2l_{\rm tube}/3$1, $l_m=2l_{\rm tube}/3$2, $l_m=2l_{\rm tube}/3$3, with $l_m=2l_{\rm tube}/3$4 end-to-end success; GR00T N1 and OpenVLA-OFT also reported $l_m=2l_{\rm tube}/3$5 end-to-end success (Haworth et al., 23 Oct 2025). Goal conditioning by point labels reduced insertion error from $l_m=2l_{\rm tube}/3$6 to $l_m=2l_{\rm tube}/3$7 for ACT and from $l_m=2l_{\rm tube}/3$8 to $l_m=2l_{\rm tube}/3$9 for $(\theta)$00, corresponding to a 59% to 74% relative reduction in $(\theta)$01 (Haworth et al., 23 Oct 2025). A plausible implication is that explicit spatial conditioning is more important for targeting precision than purely task-level language prompts.

End-to-end autonomy claims elsewhere remain task-specific. µSTAR reported the first instance of a robotic system autonomously performing vascular anastomosis on real tissue, with 90% of stitches placed without human backup, bubble leak pressure $(\theta)$02, lumen reduction $(\theta)$03, bite-depth $(\theta)$04, suture-spacing $(\theta)$05, and time per stitch $(\theta)$06 (Haworth et al., 2024). At the same time, the paper notes that knot tying was still performed manually and that tissue slippage could occur because nitinol holder grip force $(\theta)$07 was sometimes insufficient under puncture loads of $(\theta)$08 (Haworth et al., 2024).

Recurring limitations across the SutureBot literature are unusually consistent. The magnetic design is planar only and notes that full 3D freedom is required for in vivo tasks, that deeper actuation beyond 5 cm may need higher coil power and advanced cooling, and that the current 12 G diameter is large for some MIS accesses (Erin et al., 2021). Thread-grasp pipelines report that self-intersecting thread shapes are not yet supported and propose extensions such as DLO-specific graph searches and a transverse-retry action (Joglekar et al., 2024). The end-to-end benchmark identifies limited trial counts, degraded performance under unseen lighting, toolsets, and wound geometries, lack of temporal context during deep tissue penetration, and dependence on manual goal selection through a GUI (Haworth et al., 23 Oct 2025). Constraint-aware needle tracking proposes adding a fifth “compliance” dimension to model fingertip motion under tissue drag, as well as joint reasoning over multiple needles or instruments (Chiu et al., 2022). Collectively, these limitations indicate that SutureBot remains a research program centered on integrating precision perception, physically valid state estimation, reliable tissue interaction, and long-horizon policy execution, rather than a solved clinical technology.