Haptic Shoulder

Robots that reposition, dress, or rescue people have to move human limbs, but developing those behaviors means testing them on humans, which is slow, risky, and hard to repeat. Medical training has an answer for this: the phantom, a stand-in body you can practice on. We built the robotics equivalent for the shoulder.

Why the shoulder is hard

The shoulder isn’t a simple ball joint. Its reachable range depends on the configuration you’re already in, how far you can rotate the arm changes with how far it’s raised, and the limits are coupled across axes. A device that renders fixed per-axis limits doesn’t feel like a shoulder.

Shoulder joint limits are coupled: the reachable range on one axis depends on where the others are.

The human shoulder's reachable configuration space.

The device's mechanical range exceeds the human one, so the rendered limits, not the mechanism, define the boundary.

The rendered range of motion shifts with humeral rotation, reproducing the coupling that makes a real shoulder feel like a shoulder.

The mechanism

An inverted U-joint arrangement gives the device its anatomically-shaped workspace, with one motor-encoder module per anatomical degree of freedom, so actuator space maps directly onto shoulder joint space, and there’s no kinematic solving between what the motors do and what the joint does.

The inverted U-joint places all rotation axes through a single joint center that matches the human shoulder's, in CAD and in the built device.

Testing robots on it

A robot manipulating the phantom arm, pHRI planning tested and deployed on real hardware without a human subject in the loop.

Interface force through one assisted-motion trial, with and without the device rendering joint limits. Peak force rises only from 10.17 N to 10.84 N, about 7%. That small gap is the useful result: rendering realistic limits does not make the interaction meaningfully harsher, so the phantom can be used to develop and tune a policy without the added forces themselves becoming a confound.

Published at ICRA 2025 (co-first author). Provisional patent in progress.