Why Deformables Are the Hard Case
A rigid object has six degrees of freedom, so a policy can learn its dynamics from modest data. A towel has, for practical purposes, infinitely many. Cloth self-occludes, so the state you need is hidden behind the state you can see. Simulation of fabric at training fidelity remains expensive and unfaithful, which strangles the sim-to-real shortcut that rescues rigid-object tasks. The result is that deformable manipulation depends on real demonstrations more heavily than any other category, at exactly the moment when real demonstrations are hardest to find.
Open corpora reflect this. The large public datasets are dominated by rigid tabletop work, and their deformable content clusters around one entry-level task: folding. If your roadmap includes laundry beyond folding, cable routing, packaging, bed making, or any garment operation with tools, you will be collecting your own data. The honest breakdown of what open data does cover is in our custom versus open comparison.
What Good Deformable Data Looks Like
Deformable datasets fail in characteristic ways: every episode starts from a neatly flattened item, one fabric stands in for all fabrics, and the demonstrations contain no recoveries. A dataset built to train a deployable policy needs the opposite:
- Initial-state diversity, on purpose. Crumpled, twisted, inside-out, draped over an edge. The initial-state distribution is specified and staged, episode by episode, because a policy that has only seen flat cloth cannot un-crumple its first real towel.
- Material coverage. Weight, stretch, stiffness, and friction change fabric behavior more than geometry does. The spec names the material set: denim behaves nothing like chiffon, and knits behave nothing like wovens.
- Recovery demonstrations. Grasps slip on cloth constantly. Demonstrating realistic failure states and the way out of them is what separates data that trains robust policies from data that trains demo-day policies.
- Views that fight occlusion. Wrist cameras plus scene cameras with placement chosen per task, so the delivered episodes carry enough viewpoints to make the hidden state learnable.
- Success criteria that survive ambiguity. "Folded" admits degrees. We write measurable criteria into the spec, with staged references, so labels mean the same thing in episode 40 and episode 4,000.
The Category Is Wider Than Cloth
Cloth gets the headlines, but the deformable family covers most of what industrial and household robots will actually touch. Cables and wire harnesses combine deformation with tight insertion tolerances, and harness assembly remains one of the most stubbornly manual jobs in electronics and automotive production. Flexible packaging, bags, pouches, and films, dominates logistics picking, where every item is a slightly different shape by the time it reaches the gripper. Food handling adds compliance limits: the demonstration has to show a grasp that holds the item without crushing it. Each of these subcategories has its own diversity axes and its own success criteria, and each is specced separately rather than lumped under one label.
Labels for deformables need more thought than rigid-object labels. Success is often a distribution over acceptable end states rather than a pose tolerance, and intermediate states matter: a policy that learns to flatten before folding needs the flattening stage labeled. We write stage taxonomies into the spec and annotate against them during collection, when the operator's intent is known, instead of guessing from video afterward.
How We Collect It
Deformable work runs on the same operation as the rest of our collections: teleoperated rigs for on-robot state-action pairs, UMI-style handheld grippers for volume, and instrumented egocentric capture where human dexterity is the point. Purpose-built capture studios in Asia Pacific keep lighting, camera geometry, and material libraries constant across sessions, and every episode passes per-episode alignment and completeness checks before delivery in LeRobot, RLDS, or GR00T-compatible format.
Garments are the deep end of this category and our specialty: skilled operations such as sewing, ironing, fastening, hanging, and dressing forms, demonstrated by trained garment operators. That work has its own page, and the underlying automation domain is covered in our textile and garment robotics guide.
Name the objects and the operations. We reply with a coverage plan across materials and initial states, an episode estimate, and a 60-day delivery date. Get a collection quote.