njb15/corl2026-suturebot-act

SutureBot ACT policy for the CoRL 2026 tutorial

The saved ACT checkpoint used in Segments 3 and 4 of Robot Learning on the Path to Surgical Autonomy, shared with the SutureBot authors' permission. Research and educational purposes only; not for clinical use.

Files and loading

• dot_best_6311.ckpt: epoch 6311, containing model_state_dict and epoch only. Optimizer and scheduler states were removed from the training checkpoint. The weights retain their original FP32 precision.
• config.json: model architecture, fixed DistilBERT task embeddings, parameter counts, preprocessing conventions, checkpoint hashes, and provenance.
• checkpoint_normalization.json: means and standard deviations for position/jaw channels; rotation channels pass through.
• LICENSE: the existing authors' permission notice.

The checkpoint, normalization, and license are unchanged from model repository commit a6ece619d7eee45ce85c0561066643cadc526e68. In the original repository, the upload script branches from that commit and uploads only this card and config.json. With --namespace, it publishes the unchanged checkpoint, statistics, and permission notice to the new model repository instead. The original commit remains source provenance.

The participant notebook is OpenHTutorial_SutureBotACT_Inference.ipynb. It loads weights with torch.load(..., map_location="cpu", weights_only=True, mmap=True) and strictly loads all model-state tensors. Pin the model and dataset revisions to their reviewed release tags or commit hashes. The intended release tag is corl2026-v1; local staging does not create it.

Inputs and outputs

The policy uses left-endoscope and both wrist RGB images resized to 480×360, task embeddings for needle pickup / needle throw / knot tying, and zero robot-state inputs. Needle throw uses blue insertion and green exit dots over the endoscope image. The image/action policy has 122,402,525 parameters; the frozen instruction encoder has 66,362,880 parameters and does not run for each chunk.

Each forward predicts 60 actions at 30 Hz, with 20 values per action: per arm, translation offset, six relative-rotation values, and jaw angle. The notebook reverses normalization and converts predictions to absolute poses using the measured starting pose. Its released deployment convention subtracts the first predicted translation.

Related cached data and compatibility

Segment 3 goals, metadata, and predictions belong in the separate dataset repository njb15/corl2026-seg3-cache. This link is the intended upload destination; local staging does not publish it.

Existing predictions/act_1441.npz through act_1451.npz and predictions/summary.json remain in the historical model repository for Segment 4 compatibility. They contain absolute 16-D poses under open_loop, chunk, and dream, using a 50-step receding-horizon protocol. The upload script does not replace or delete them. Segment 3 uses its own per-episode files in the dataset repository, with a different schema and goal-conditioned inputs.

Limitations, permission, and sources

The checkpoint's training split was not recorded and may overlap the OpenH evaluation clips. Treat offline comparisons as teaching demonstrations, not benchmark scores or physical task-success rates. The old Segment 4 needle-throw caches were generated without goal dots; the Segment 3 cache uses matched source annotations.

The authors' permission is recorded in LICENSE; no general-purpose license has been attached by them. Contact the authors for other uses. Source dataset terms also apply to data obtained from the original sources.

• Code: SutureBot ACT, pinned source.
• Haworth et al., SutureBot: A Precision Framework & Benchmark For Autonomous End-to-End Suturing, NeurIPS 2025 Datasets and Benchmarks (arXiv:2510.20965).
• Open-H-Embodiment dataset and report.