Researchers at The Hong Kong University of Science and Technology (HKUST) have built an embodied sensorimotor robotic arm called EmArm that can feel touch along its entire arm and react in real time. The project points to a shift from camera‑centric robots toward embodied, multimodal perception.
Soft optical skin embedded in the arm
EmArm is inspired by the human “bone–muscle–skin” structure. Instead of adding patches of electronic skin on top, the team deeply embedded a large‑area soft optical tactile skin into a rigid robotic skeleton:
- more than 71,000 mm² of soft skin covers the upper and lower arm;
- sub‑millimeter internal cameras track deformations and force distribution;
- spatial resolution reaches 0.085 mm, comparable to human fingertip skin;
- two‑point discrimination threshold is as low as 18.2 mm, better than the typical 30–50 mm of human arm skin.
Because sensing is optical, EmArm avoids dense wiring and is inherently resistant to electromagnetic, light, temperature, humidity, and acoustic interference. Experiments show up to 97.9% tactile sensing fidelity under five types of environmental disturbance.
Crucially, the embedded skin does not compromise mechanics: EmArm maintains repeat positioning accuracy similar to conventional industrial arms, withstands >15 kg normal load and 7.35 Nm torsion, and still delivers high‑precision actuation.
Perception–action loop and “emotional” touch
To turn EmArm from a rich sensor into an intelligent agent, Prof. Yajing Shen’s team designed a perception–action closed loop. The framework fuses joint‑angle data (proprioception) with tactile input (exteroception) in real time and extracts features such as contact location, area, force, and number of contact points. These features are fed directly into the motion controller.
A demo with a basketball placed off‑center on the arm shows EmArm comparing real‑time tactile features to a pre‑learned policy and continuously adjusting joint angles and contact forces to keep the ball balanced.
In human–robot interaction tests, EmArm recognizes five touch types — stroking, pinching, poking, tickling, and hitting — and maps them to five emotional feedback expressions: happiness, sadness, curiosity, shyness, and anger. Reported recognition accuracy reaches 99.8%.
In manipulation and safety scenarios, EmArm uses whole‑arm tactile feedback to stabilize off‑center loads, perform trajectory replanning when unseen dynamic obstacles (like a human hand) enter its path, trigger emergency stops, classify obstacles as temporary or permanent, and autonomously re‑route — even when vision is blocked.
The team sees EmArm as a step from “externally attached perception” to “intrinsic intelligence”, and from vision‑only systems to multimodal fusion. The same embodied sensorimotor architecture could extend to robotic legs, torsos and full humanoid platforms for home service, industry and medical assistance, and help drive advances in embodied AI — AI with a physical body that learns directly from the real world.
Source: The Hong Kong University of Science and Technology, Phys.org
