Session: SYMP S-5: Mechano-Intelligent Robots

Paper Number: 140641

140641 - Next-Generation Exploratory Robots Through Physical Intelligence 

Exploratory robots are indispensable in space exploration, deep-sea discovery, and rescue missions. In these challenging environments, exploratory robots use onboard electronic sensors such as cameras, IMU, and more recently, electronic skins to sense and navigate without human intervention. However, their complete reliance on electronic systems makes them vulnerable to extreme conditions such as extreme temperatures, high pressure, and electromagnetic interference. Additionally, the complexity and fragility of these systems drive up costs and restrict deployment. Therefore, there is an urgent need for alternative solutions that overcome the limitations of traditional electronics-based systems.

The recent advent of physical intelligence, inspired by biological organisms, offers a novel solution to the challenges faced by exploratory robots. With a strategic use of structures and materials, physical intelligence can integrate sensing, actuation, adaptation, and even computational functions into the body of a robot, which will dramatically enhance its overall capabilities, resilience, and robustness.

In this talk, we will introduce a prototype robotic system that can sense the surrounding environment not only from electronic sensors but also through its physical bodies. This system incorporates a body made from metamaterials, referred to as the Physical Network (PN), that is completely passive and can be manufactured from strong and resilient materials. Any circumferential interactions such as forces and impacts will generate mechanical waves that propagate through the PN to a single accelerometer—the only electronic sensor in the system. A Neural Network (NN) will be trained to detect various circumferential interactions based on the time-series data from the accelerometer. The system will be able to differentiate the type, amplitude, and pattern of circumferential interactions, as well as assess the mechanical properties of contacting objects, all through a single electronic sensor at its center—a level of sensing capability currently achieved by dense sensor arrays.

The project's primary challenge lies in the intricate co-design of the PN's geometry and the tightly-coupled computational model NN in an automatic data-driven manner. As the PN geometry directly influences the accelerometer's readings and, consequently, the NN's predictions, the PN (described by parameter set ) and NN (with parameter set ) must be jointly optimized to achieve optimal sensing capabilities. In order to achieve that, we employ a differentiable simulator that provides not only accelerometer signal A(t) but also the gradients relative to the design parameters, through which the gradient of prediction loss can be backpropagated all the way from NN to PN (see figure). Two networks, PN and NN, will then be updated concurrently based on their gradients to losses, leading to a PN geometry ideally suited to the NN's processing. The optimized PN geometry will be manufactured and hardware experiments will be conducted to fine tune the NN to bridge the simulation-to-reality discrepancies.

With its only electronic component shielded by the metamaterial casing, the proposed system exhibits an enhanced resistance to severe conditions, and powerful tactile sensing capabilities that can advance the field of soft robotics, robotic manipulation, wearable robots, and human-robot interaction. Furthermore, the design's simplicity not only reduces manufacturing costs but also makes the technology more accessible for broader application in exploration. More importantly, this project will usher in a new paradigm in robotics where the robot's body (PN) is no longer a passive structure but an active participant, intelligently trained in tandem with its computational brain (NN) for specific tasks.

Presenting Author: Bolei Deng Georgia Institute of Technology

Presenting Author Biography: Bolei Deng is an Assistant Professor at the Guggenheim School of Aerospace Engineering at the Georgia Institute of Technology. Graduating with a B.S. in Engineering Mechanics from Zhejiang University in 2016, he later earned his Ph.D. in Mechanical Engineering and Material Science from Harvard University in 2021, under the guidance of Prof. Katia Bertoldi. Following this, he undertook a joint postdoctoral position at MIT between the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Department of Mechanical Engineering.

His research primarily focuses on employing artificial intelligence for the design and optimization of mechanical metamaterials across various scales. He has a keen interest in understanding and leveraging nonlinear behaviors, including nonlinear dynamics, multistabilities, and fracture. His work spans from developing ultra-strong and tough metamaterials to innovations in robotics, mechanical computing, and physical intelligence.

Authors: