Skip to content
Bud's Wiki

Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge

TL;DR

Section titled “TL;DR”

Sharpa (co-organizing the IROS 2026 Robotic Origami Challenge with @chris_j_paxton, @micoolcho, @DJiafei, and @lightwheel) has released the Origami Dataset — the training half of a challenge that asks teams to fold a six-fold traditional Japanese paper airplane on a bimanual rig with two 22-DoF SharpaWave five-finger tactile hands. The dataset contains 682 episodes / 51 collection sessions / 4.76M frames at 30 FPS across 6 synchronized camera views (head + wrist), 10 fingertips × 6-axis force/torque signals synchronized frame-by-frame with vision and 65-D joint-space state/action, and ships with two tactile video formats (raw sensor stream + lighter deformation map) plus full LeRobot 3.0 coverage. Positions itself as “one of the largest tactile datasets out there” and is deliberately targeted at the pretraining bottleneck for tactile-sensing manipulation policies.

Key claims

Section titled “Key claims”
  • Dataset scale: 682 episodes, 51 collection sessions, 4.76M frames at 30 FPS on paper folding [tweet body].
  • Modalities: 6 synchronized camera views (head + wrist), 10 fingertips × 6-axis force/torque signals synchronized frame-by-frame with vision and motion, 65-D joint-space state and action across two arms + two dexterous hands [tweet body].
  • Two tactile video formats are shipped: a raw sensor stream (maximum fidelity) and a lighter deformation map derivative — mirroring the deform-map abstraction Sharpa’s own Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map simulator commits to [tweet body].
  • Sharpa argues the dataset is uniquely valuable because (a) contact is registered the instant the finger touches the paper (low tactile latency) and (b) the sensor is sensitive enough to register the tiny contact forces produced by thin paper that “most tactile systems either miss or drown in noise” [reply tweet 2082756234818818307].
  • Full LeRobot 3.0 coverage across all seasons — drop-in for existing long-horizon imitation-learning and visual-tactile representation-learning pipelines [tweet body].
  • The task the dataset targets is folding a kami hikōki traditional paper airplane in exactly six folds, judged by a Nippon Origami Association Grand Master on crease accuracy, structural fidelity, symmetry, and paper integrity [linked challenge page].
  • Registered teams also receive 500+ episodes of teleop demos on the target folding sequence, an NVIDIA Isaac Sim high-fidelity sim with paper physics, and remote-lab access to the same rig that will be used at IROS in Pittsburgh [linked challenge page].
  • Companion release: SharpaWave URDF + tactile assets are open-sourced at sharpa.com/pages/open-source [tweet body].

Method

Section titled “Method”

The Origami Dataset was collected by teleoperating a bimanual Sharpa rig — two arms each carrying a 22-DoF SharpaWave five-finger tactile hand — through the traditional six-fold Japanese paper airplane sequence. Each frame packages six camera views (head and wrist cameras from both sides), a 65-D joint-space snapshot of both arms and hands, and 10 fingertip 6-axis F/T channels synchronized to the same 30 FPS clock as the vision streams. The tactile channel is exported in two forms: the raw sensor stream (for fidelity) and a per-fingertip deformation map (for compactness) — deliberately the same normal-projection deform-map representation Sharpa argues for as a sim-to-real observation-space contract in Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map. The whole dataset is repackaged in LeRobot 3.0 layout so existing imitation-learning stacks consume it without adapters.

The associated challenge closes the loop: teams train on the dataset (and optional Isaac Sim rollouts), upload policies to Sharpa’s remote lab for identical-hardware evaluation, and are ranked by an Origami Grand Master’s judgment on the folded airplane — a rubric that is deliberately outside the ML metric-hacking cycle.

Results

Section titled “Results”

Not applicable — this is a dataset + benchmark release, not a paper reporting policy results. The tweet does not claim any policy numbers on the dataset; the ranking will be established by the participating teams’ submissions at IROS 2026.

Why it’s interesting

Section titled “Why it’s interesting”

Fills the exact gap Data Pyramid for Embodied Manipulation named as the first of six open challenges for embodied manipulation — large-scale tactile datasets — with a bimanual, whole-hand, force/torque-synchronized corpus of a genuinely delicate task. It also operationalizes Sharpa’s own Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map observation-space bet (deform-map as the canonical tactile channel) at dataset scale rather than sim scale, and stakes out a different placement-of-taxels position than Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks: 10 fingertip 6-axis F/T sensors, no palm or proximal-phalanx taxels — the anthropomorphic-fingertip route rather than whole-hand coverage. Paper folding is a strong stress test for the Tactile-Reactive Dexterous Hand: High-Frequency Physical Interaction “kHz-class touch dominates” thesis: paper is thin, deformable, easy to occlude, and unforgiving of contact timing in a way rigid-object tasks are not.

See also

Section titled “See also”