HKUST publishes study of EmArm robot arm with whole-arm tactile sensing in Nature Sensors

A research team at the Hong Kong University of Science and Technology published a study of EmArm, a robot arm with tactile sensing across its entire arm, in the international journal Nature Sensors. The paper was published online on July 20, 2026, and appeared in the August issue, leaderobot.com reported. The soft tactile skin covering the upper and lower arms spans more than 71,000 square millimeters, and sensing fidelity was 97.9% under five types of environmental interference.

The research team demonstrated in experiments how tactile information from the entire arm could be used to adjust an object's posture and to stop or replan a route around dynamic obstacles that could not be detected visually. The design feeds tactile information back into control rather than limiting it to contact detection, making it relevant to the study of robot arm operation in situations that are difficult to assess through vision alone.

A key feature of EmArm is not only the skin's area, but also its configuration: cameras are placed inside the mechanical skeleton to read skin deformation as visual information. Rather than wiring numerous tactile elements individually, the system extracts contact location, area, force and number of contact points from deformation caused by contact. The study's novelty, therefore, lies not simply in covering a large area of the arm, but in integrating the mechanism and processing needed to convert broad-area tactile information into a form that can be used for control. The research team presented a framework that combines extracted tactile features with joint angles and feeds them back into motion control. In a basketball posture-adjustment demonstration, the arm changed the angles of its upper and lower sections to stabilize an object whose center of gravity was off balance. In an obstacle experiment, it detected a dynamic obstacle that the vision system could not see, then stopped and replanned its route. These demonstrations show a design that links touch to motion correction rather than treating it merely as a record of contact. However, the functions demonstrated in the experiments do not establish long-term operation in real-world environments or versatility across a wide range of objects. The performance figures also need to be read in context. The 0.085-millimeter spatial resolution, 18.2-millimeter two-point discrimination threshold and 97.9% sensing fidelity under five types of interference are results reported by the research team; they do not show that the same performance would be achieved with different robots or in different operating environments. The results showing that the system withstood pressure loads exceeding 15 kilograms and a torsional load of 7.35 newton-meters, while retaining repeatability comparable to conventional robot arms, can be viewed as testing whether soft skin can coexist with rigidity and drive performance. The next issues to examine are the durability of the skin and internal cameras, replacement and maintenance, processing delays for contact information, and false detections and safety in real-world environments. The research team also discussed the possibility of future applications to legs and torsos, but the work presented so far consists of an arm prototype and experimental results. Manufacturing costs, production methods and actual deployment requirements are not clear from the announcement, and whether expanded tactile coverage offers operational benefits needs to be assessed alongside these factors.

EmArm has soft optical tactile skin on the outside of a rigid mechanical skeleton, with submillimeter cameras embedded inside. The cameras capture deformation of markers on the inner surface of the skin, and the system extracts contact location, area, force and number of contact points. It uses a perception-action framework that combines this information with joint angles and feeds it back into motion control. The research team demonstrated experiments identifying different types of human touch and responding with facial expressions, adjusting the posture of a basketball resting on the arm with its center of gravity off balance, and stopping and replanning a route around dynamic obstacles that could not be detected visually. The team reported 99.8% accuracy in identifying types of touch. According to the team, the skin maintained repeatability comparable to conventional robot arms and withstood pressure loads exceeding 15 kilograms and a torsional load of 7.35 newton-meters.