Research map / Vision-language and multimodal / Robot action and navigation
SutureBot: A Precision Framework & Benchmark For Autonomous End-to-End Suturing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SutureBot, a dVRK suturing benchmark with a 1,890-demo dataset and a goal-conditioned (insertion/exit point) low-level VLA/ACT framework with a high-level task-prediction policy, evaluated on end-to-end suturing with precision measured via UV markers.
Problem
Autonomous end-to-end robotic suturing requires long-horizon dexterous manipulation with precise needle insertion, but no public real-world benchmark, large dataset, or precision metrics exist, and no fully autonomous suturing has been shown on physical hardware.
Approach
Condition low-level imitation policies on explicit goal inputs (insertion and exit points marked on the endoscope image, plus masks or distance maps) so policies learn to reach user-specified suture targets, with a Swin-T high-level policy sequencing language-conditioned subtasks.
Key results
- Goal-conditioned point-label ACT achieved 1.3 mm insertion error vs 3.2 mm with no goal; pi0 with point labels achieved 1.0 mm vs 3.9 mm without goal.
- Multitask ACT was the best low-level policy with 9/10 pickup, 8/10 throw, 9/10 knot tie and 3/10 end-to-end suturing; OpenVLA-OFT achieved 0 success on all subtasks.
- High-level task predictor reached 88.73% accuracy and F1 0.92 offline; ACT with high-level policy performed comparably to the oracle variant.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| condition needle placement on point labels rather than no goal (ablation) | ACT · SutureBot needle throw | insertion error (mm): 3.2±2.2 -> 1.3±0.9 | Helped |
| condition needle placement on point labels rather than a distance map (ablation) | ACT · SutureBot needle throw | insertion error (mm): 2.6±1.5 -> 1.3±0.9 | Helped |
| condition needle placement on point labels rather than a binary mask (ablation) | ACT · SutureBot needle throw | insertion error (mm): 2.9±1.7 -> 1.3±0.9 | Helped |
| condition needle placement on point labels rather than no goal (ablation) | π₀ · SutureBot needle throw | insertion error (mm): 3.9±2.5 -> 1.0±1.3 | Helped |
| condition needle placement on point labels rather than a distance map (ablation) | π₀ · SutureBot needle throw | insertion error (mm): 2.1±1.1 -> 1.0±1.3 | Helped |
| condition needle placement on point labels rather than a binary mask (ablation) | π₀ · SutureBot needle throw | insertion error (mm): 1.8±1.2 -> 1.0±1.3 | Helped |
| use ACT instead of GR00T N1 | low-level policy · SutureBot | throw success: 2/10 -> 8/10 | Helped |
| use ACT instead of π₀ | low-level policy · SutureBot | end-to-end sutures: 0/10 -> 3/10 | Helped |
| replace high-level task prediction with human-provided language conditions | ACT · SutureBot | end-to-end sutures: 3/10 -> 2/10 | No clear effect |
| post-train the π₀ checkpoint on SRT-H cholecystectomy trajectories (ablation) | π₀ · SutureBot | insertion error (mm): 1.9±1.0 -> 2.2±0.7 | Hurt |
| initialize π₀ from scratch with a PaliGemma backbone instead of the standard checkpoint (ablation) | π₀ · SutureBot | insertion error (mm): 1.9±1.0 -> 3.7±3.6 | Hurt |
| evaluate on alternate lighting | ACT · SutureBot | pickup success: 9/10 -> 3/10 | Hurt |
Surprising findings
- OpenVLA-OFT failed all subtasks (0/10) and GR00T N1 succeeded in only 1/10 pickups, despite strong general benchmark performance.
- Pretraining on 20,000 cholecystectomy trajectories (pi0 Chole) did not clearly improve in-domain suturing performance over the standard pi0 checkpoint.
Limitations
- Only 10 trials per experiment due to time constraints, limiting statistical power for subtle effects.
- Performance drops under lighting and tool changes; end-to-end autonomy achieved only by ACT in 3/10 trials.
- Relies on manual target-point annotation and lacks historical context; high-level policy evaluated only offline and in oracle settings.
Experimental setup
Datasets: SutureBot dataset (1,890 dVRK demos) SRT-H cholecystectomy trajectories
Benchmarks: SutureBot end-to-end suturing on dVRK suture pad
Metrics: task success rate insertion error (mm) exit error (mm) procedure time (s) end-to-end success
Baselines: ACT pi0 GR00T N1 OpenVLA-OFT SRT-H pi0 Chole No-goal conditioning
Builds on: ACT SRT-H OpenVLA-OFT pi0 DAgger
Closest papers by meaning
- SurgicAI: A Hierarchical Platform for Fine-Grained Surgical Policy Learning and Benchmarking (NeurIPS 2024 · similarity 0.87)
- HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing (NeurIPS 2026 · similarity 0.86)
- ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · similarity 0.86)
- Scaffolding Dexterous Manipulation with Vision-Language Models (NeurIPS 2025 · similarity 0.84)
- RobotSmith: Generative Robotic Tool Design for Acquisition of Complex Manipulation Skills (NeurIPS 2025 · similarity 0.84)
- UniGraspTransformer: Simplified Policy Distillation for Scalable Dexterous Robotic Grasping (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.