Garment folding teleoperation capture

Garment folding teleoperation capture is custom data collection for robots that must grasp cloth, make folds, and recover when layers slip or bunch. Operant records an operator driving your robot through those contacts, with synchronized RGB-D, proprioception, fold-state labels, and recovery metadata. The episodes match your gripper, workspace, and cloth set, so a folding policy can train on the deformations it will meet in deployment.

What we collect

Folding data is valuable because the object changes shape under the grasp. A capture plan can include pick-up, shake or flatten, corner grasp, fold, stack, and reset. Variation can cover shirts, towels, pants, thickness, color, wrinkles, damp cloth, patterned fabric, and starting poses that hide a corner.

Successful folds are only part of the program. Slipped layers, double grasps, bunched hems, corners that fall open, and the recovery that follows can be captured on purpose. Those episodes are what let evaluation reflect cloth behavior outside a single tidy demo.

This scenario is about deformable garments. Rigid pick work stays on pages such as bin picking teleoperation. Kitchen object tasks stay on dual-arm kitchen manipulation.

Sensors and modalities

Typical capture includes calibrated multi-view RGB-D, a wrist or gripper camera aimed at the cloth contact, proprioception, gripper or hand state, and operator commands. Force/torque is included when the grasp or the fold depends on tension. Operant aligns these streams through the multi-sensor synchronization service so the image of a slipping layer and the command that corrects it refer to the same moment.

The rig is scoped to your platform. Some teams need a close view of the hem. Others need a full-torso view of a bimanual fold. The delivery is the observation set, action space, and metadata your training pipeline can consume.

How capture works

A pilot validates the teleoperation mapping, camera views, episode boundaries, and fold labels on a small garment set. Operators then capture demonstrations through the teleoperation capture service. QA checks sync, calibration, metadata completeness, and whether the fold or the failure is actually visible.

Scale adds garment types, starting poses, and outcome classes from the statement of work. Repeating one shirt in one pose is not the goal. The goal is the cloth variation your policy has to generalize across, collected on the robot you will deploy.

QA and metadata

Each episode can carry garment type, size, color, thickness, starting pose, fold step, outcome, recovery action, operator ID, and surface notes. QA compares the files with the agreed modalities, timestamp tolerance, calibration version, and required labels.

Your team can then separate clean folds from recoveries, or hold out a garment type that never appeared in training. Labels follow your schema and are applied the same way across the program, consistent with robot demonstration data.

Who it is for

This scenario is for teams teaching a robot to fold or straighten cloth when simulation and rigid-object datasets miss layer slip, occlusion, and recovery. It fits humanoid robotics programs and other manipulation stacks that share the same gripper and action space.

To scope garment-folding capture around your robot and cloth set, book a discovery call.