Sensing Through Mechanical Metamaterials

Exploratory robots play a crucial role in space, deep-sea, and rescue missions. However, their reliance on electronic sensors such as cameras, IMUs, and electronic skins makes them susceptible to extreme temperatures, high pressure, and electromagnetic interference. These vulnerabilities, coupled with the complexity and cost of electronics, limit the deployment of traditional systems. There is a need for alternative sensing methods that overcome these challenges.

Inspired by biological organisms, physical intelligence offers a promising approach by integrating sensing, actuation, adaptation, and computation directly into a robot’s structure. This talk introduces a novel robotic system that senses the environment through its physical body. The system features a metamaterial-based structure, the Physical Network (PN), which is entirely passive yet highly resilient. When subjected to external forces, mechanical waves propagate through the PN to a central accelerometer—the sole electronic sensor in the system. A Neural Network (NN) processes the accelerometer data to classify the type, magnitude, and pattern of interactions, as well as determine the mechanical properties of contacting objects, matching the capabilities of traditional dense sensor arrays.

The system's design hinges on the co-optimization of the PN’s geometry and the NN’s model. Using a differentiable simulator, we capture accelerometer signals and compute gradients with respect to the design parameters, allowing for simultaneous updates to the PN and NN based on performance. This results in an optimized PN tailored to the NN’s processing requirements. The design is then fabricated and fine-tuned through hardware experiments to address simulation-to-reality discrepancies.

Shielded by its metamaterial structure, this system exhibits enhanced resilience in harsh conditions and offers powerful tactile sensing for applications in soft robotics, robotic manipulation, and wearable technology. The project reimagines robotics, transforming the body from a passive structure into an active, co-trained component alongside its computational model.