Robotic Origami Challenge: Fold Plane Demonstrations

Bimanual paper-folding dataset: 682 LeRobot episodes with vision and touch

License

CC-BY-4.0

Overview

Robotic Origami Challenge: Fold Plane Demonstrations is a real-world teleoperation dataset for folding a traditional paper airplane with a bimanual dexterous robot. Released by Sharpa in LeRobot format for the IROS 2026 Robotic Origami Challenge, it holds 682 episodes and 4,763,267 frames captured at 30 FPS across 51 collection seasons. Every frame carries six synchronised video streams — head left and right, wrist left and right, tactile deformation and raw tactile — alongside 65-dimensional joint state and action and 60 tactile channels read from ten fingertips. Released under CC-BY-4.0 in LeRobot v3.0 format, with a v2.1 export.

Origami is an unusually hard manipulation benchmark: paper is thin, deformable, and creases irreversibly, so a policy has to coordinate two hands precisely while reading contact it cannot see. This dataset captures human teleoperation of exactly that task on Sharpa's bimanual dexterous rig — the same arms-plus-Sharpa-Hands configuration used in the challenge's simulator, its remote evaluation lab, and on the IROS floor. The target is a traditional Japanese paper airplane folded from a 15 x 15 cm sheet of at least 60 gsm in six folds, scored on crease accuracy, structural fidelity, symmetry and paper integrity. Data ships in LeRobot v3.0 format with a v2.1 export, so it drops into existing imitation-learning pipelines without conversion.

Why it matters

Open manipulation data is overwhelmingly rigid-object pick-and-place recorded from cameras alone. This set is deformable-object, bimanual, dexterous and tactile-instrumented at once — 60 force/torque channels across ten fingertips alongside six camera views — which is the combination the field has least of. Because it is CC-BY-4.0 and LeRobot-native, it is usable for training and benchmarking generally, not only by teams entering the competition it was collected for.

Problem it solves

Deformable-object manipulation lacks open, tactile-rich bimanual data; most released datasets are single-arm, rigid-object and vision-only, leaving contact-sensitive tasks such as folding without a training corpus.

Key takeaways

• 682 episodes and 4,763,267 frames at 30 FPS, spread across 51 collection seasons
• Six synchronised video streams: head left and right, wrist left and right, tactile deformation and raw tactile
• 65-dimensional joint state and action spanning two arms, Sharpa Hands and torso joints
• 60 tactile channels — ten fingertips by six force/torque axes
• CC-BY-4.0, distributed in LeRobot v3.0 format with a v2.1 export

Who it is for

• Dexterous manipulation researchers
• Deformable-object manipulation teams
• Tactile sensing researchers
• Robotic Origami Challenge entrants

Limitations

• A single task — one paper-airplane fold — on a single hardware rig, so cross-task and cross-embodiment transfer is untested
• Recorded exclusively on Sharpa Hands; the 60-channel tactile layout does not map directly onto other end-effectors
• Download sits behind a click-through agreement on Hugging Face, though access is granted automatically
• Teleoperated human demonstrations rather than optimal or autonomous trajectories
• Only the dataset is open — the surrounding challenge stack (Lightwheel simulator, remote evaluation lab) is not