HKUST's EmArm embeds more than 71,000 mm² of camera-read tactile skin in a robot arm

More than 71,000 square millimeters of soft optical tactile skin cover the upper and lower arms of EmArm, a robotic arm HKUST announced on Oct 16. The work appeared in Nature Sensors on July 20.

HKUST's claim is that whole-arm touch does not have to cost an arm its strength. The release reports repeatability comparable to conventional robotic arms, pressure loads above 15 kg and 7.35 Nm of torsion, so the skin could serve as the arm's safety sensor when cameras are blocked. The weak point is scale. Every result comes from the team's own tests on one arm, and moving the skin onto legs and a torso, which HKUST pitches as a route for humanoids and embodied AI, is still a plan.

Cameras inside the skeleton read the skin

The design copies what HKUST calls the human "bone-muscle-skin" structure. The soft skin sits inside the arm's rigid skeleton, and sub-millimeter cameras inside the arm track how force spreads across it. Because the sensing is optical, HKUST says, the arm avoids the wiring that electronic skins need and resists electromagnetic, light, temperature, humidity and acoustic interference, holding sensing fidelity at up to 97.9% under those five conditions. It reports 0.085 mm spatial resolution, which it compares with human fingertip skin, and a two-point discrimination threshold as low as 18.2 mm, against the 30 to 50 mm typical of human arm skin.

Touch features go straight into motion control

The control loop merges joint angles (proprioception) with touch (exteroception) in real time. It extracts contact location, area, force and the number of contact points and passes them to motion control. With a basketball resting off-center on the arm, HKUST says, EmArm matches live tactile features against a pre-learned policy and issues joint-angle commands to keep the ball stable. Prof. Shen Yajing, the corresponding author, with Dr. Tang Yifeng as first author, explains why the team did not train the arm end to end: that approach "will not only require massive annotated data but also demand high computational power, making it difficult to run in real time on a robot."

In a touch-interaction test the arm told apart five behaviors (stroking, pinching, poking, tickling and hitting) with up to 99.8% accuracy and answered each with an emotional expression. In a replanning test, a human hand that vision could not detect entered its path. The arm felt it, stopped, judged whether the obstacle was temporary or permanent, and planned a way around.

The two headline numbers measure different things. The 0.085 mm figure describes how precisely the skin locates a single contact; the 18.2 mm figure is how far apart two contacts must be before the arm reports two. HKUST puts the first level with a fingertip and the second ahead of human arm skin.