Electronic skin: Bridging the sensory gap in embodied AI

A form of flexible electronic skin developed by the Institute of Flexible Electronics Technology of THU in Zhejiang could help humanoid robots move beyond basic motion by giving them a sense long missing from AI training: touch.

Humanoid robots need large volumes of real-world interaction data to perform complex tasks. Yet most model training relies heavily on visual information and motion trajectories, leaving robots unable to fully perceive the physical properties of the objects they handle.

The institute's flexible tactile sensing terminal is designed to close that gap. Made with its integrated flexible-device manufacturing technology, the electronic skin is thin, stretchable and able to conform to curved surfaces such as robotic fingertips and arms.

Once attached, the sensors allow a robot to detect pressure, temperature and texture in real time. Combined with force-control algorithms, the system delivers force feedback at a frequency of 1 kilohertz and a resolution of 2 percent, enabling the robot to adjust its grip continuously rather than rely on preset movements.

The system can handle fragile objects using as little as 2 newtons of force while applying up to 100 newtons when gripping heavier items. Automatic compensation helps prevent objects from slipping, while an active warning function alerts the system when its grip becomes unstable.

Such capabilities could improve robots' ability to handle delicate or irregular objects and generate the high-precision physical-interaction data needed to train embodied AI models, according to the institute. The technology also provides a lower-cost way of collecting tactile data that visual sensors alone cannot capture.

The institute said it has deployed the electronic skin across its industrial embodied AI training center, creating a continuous pipeline for collecting data from physical interactions. The facility can produce and install up to 1,000 intelligent sensing grippers a month, supporting the larger-scale training of robots for real-world tasks.