In the lab, a robot arm now senses touch across its whole length

A robot arm at the Hong Kong University of Science and Technology can now tell a pat from a poke, and it can keep a basketball balanced on its forearm. EmArm, as the team calls it, does this through more than 71,000 square millimeters of soft tactile skin that covers its upper and lower arm. The results are published in Nature Sensors and come from laboratory experiments led by Professor Shen Yajing and his team at the Mevita Robotics Lab.

Most robots feel contact only where a sensor sits, usually at the gripper. Attaching sensors to the outside of a limb adds wiring and fragility. The HKUST team took a different route, inspired by the body's "bone-muscle-skin" structure: they placed a soft optical skin inside the rigid skeleton, and submillimeter cameras read how that skin deforms when something touches it. The optical approach also cuts down on the complex wiring that electronic skins need, and the team reports it is resistant to electromagnetic, light, temperature, humidity and acoustic interference.

The measurements are precise. EmArm reaches a spatial resolution of 0.085 millimeters, which the researchers compare to human fingertip skin, and its two-point discrimination threshold is 18.2 millimeters, where typical human arm skin manages 30 to 50 millimeters. Its sensing fidelity stayed at up to 97.9% across five kinds of environmental interference. A sensor laid over a robot's structure can weaken it, so the team also tested the mechanics: the arm withstood pressure loads above 15 kg and torsional loads of 7.35 Nm, and kept repeat positioning accuracy the researchers describe as comparable to conventional arms.

The software is what turns this skin into behaviour. The arm combines joint-angle data with touch data, extracts contact location, area, force and the number of contact points, and feeds them directly into its motion control. In the basketball test, it adjusted its posture and forces continuously to keep the ball steady. In the obstacle test, a human hand entering the path, invisible to the cameras, was detected through touch; the arm stopped, judged whether the obstruction was temporary or permanent, and replanned its trajectory. In social interactions, it recognised five kinds of touch, stroking, pinching, poking, tickling and hitting, with accuracy of up to 99.8%, and answered with expressions of happiness, sadness, curiosity, shyness or anger.

And so, concretely? For now, the change is a demonstration: a robot that keeps working safely when it cannot see what is touching it. The researchers say the approach could extend to legs, torsos and other body parts, which they argue could help humanoid robots work more safely in factories, homes and medical assistance. No timeline, product or industrial partner has been announced. The figures come from the team's own experiments, and they will need to be reproduced by other groups, in real settings, before anyone can say how far they carry.