Humanoids stop manipulating with hands alone: wearable contact patches turn human demos into whole-body pushing, sweeping and carrying

Manipulation for a humanoid no longer stops at the hands: forearms, torso and knees become contact surfaces for embracing bulky objects, sweeping items together, or pushing when the hands are occupied. Much of loco-manipulation research restricts manipulation to the hands, reserving the rest of the body for locomotion and balance and typically avoiding other body contact with the environment. On a Unitree G1, the upper-body IK variant reaches 85% success and the sparse contact tracker 80% on grasp-and-embrace; the sweep-and-stow task reaches 90.63% end-to-end success and carry-and-push reaches 90%.

The system treats contact as manipulation-centric information that transfers across embodiments: where and when contact happens describes the physical interaction a task needs, and the robot chooses which body surface realizes it. Existing pipelines that learn manipulation from humans capture and imitate end-effector or joint trajectories, and do not explicitly carry body-contact information from demonstration through policy prediction to robot execution. Two pose-based interfaces without explicit body-contact objectives serve as comparisons: Humanoid UMI, tracking only end-effector trajectories, consistently fails to support the bulky object, while Humanoid UMI++ often knocks objects outside the bin or approaches joint limits.

A wearable interface combines egocentric cameras and bimanual UMI grippers with modular contact patches built from identical force-sensing modules that cover the forearms and knees and record contact timing, intensity and location. Existing whole-body tactile skins are typically tailored to a particular robot embodiment, and existing demonstration interfaces record only end-effector motions or joint angles. Each module carries a single force-sensitive resistor, a custom board streams up to 16 channels, modules communicate over a daisy-chained RS-485 bus, and magnetic connectors snap them into patches of varying shapes and sizes.

The policy predicts end-effector trajectories together with per-sensor contact goals, and the controller tracks a contact location only when its contact state is active, leaving the remaining degrees of freedom free to choose a robot configuration. Controllers designed to track motion references or end-effector targets lack an explicit objective for achieving contact goals on the body. Replaying demonstration trajectories in simulation, adding contact goals reduces contact surface distance for every controller at little cost to end-effector tracking accuracy; the sparse contact tracker achieves the smallest contact distance but a higher end-effector error than WBIK+SONIC.

The report covers PanTact deployed on a Unitree G1 across four task families: grasp-and-embrace, sweep-and-stow, carry-and-push, and a simulation controller comparison. The intended readers are researchers and engineers working on humanoid manipulation, imitation learning and whole-body control. The robot carries no additional body contact sensors; contact goals are predicted by the policy and realized by the controller.

Contact detection binarizes force-sensitive-resistor signals against a per-channel adaptive baseline because resistance drifts with temperature and mounting pressure. The sparse contact tracker requires simulation training and depends on the training distribution, and its end-effector error is higher than WBIK+SONIC, while WBIK+SONIC's dense reference specifies postural details that are not essential to the task and its IK solver does not model the tracker's execution limits.