A Systematic Way to Train Physical Intelligence
Oral-In-person · Withdrawn
Abstract
The implementation of intelligence in mechanical systems is essential for advancing autonomous systems. Such systems must actively adapt to environments and demonstrate capabilities of learning, memory, and decision-making. The reliance on either solely mechanical platforms or computational resources limits overall performance of intelligence.
Our approach addresses these challenges through the development of physical intelligence (PI), which functions as an intelligent body interacting with and adapting to environmental stimuli, complementing computational intelligence (CI) which serves as the brain. The proposed system employs mechanical metamaterials with adaptable properties, utilizing a single central accelerometer as the sole electronic sensor. When external forces are applied to the metamaterial platform, mechanical wave propagation is detected and analyzed. Neural networks are trained to classify interaction sources or estimate properties of external objects based on these measurements.
Concurrently, the metamaterial structure undergoes optimization to enhance neural network performance through a two-tier framework: design optimization (high-level) and neural network training (low-level). The metamaterials are well known as their programmability and adaptability. Previous research proposed the effectiveness of inverse design of metamaterials to achieve target properties (e.g. maximizing energy absorption) using differentiable simulations of metamaterials. Inspired by the efficiency of differentiable simulation to optimize the designs, we implement differentiable simulation of metamaterials to obtain gradient information with respect to design parameters. Since the high-level optimization for design includes the iterative computational graph of training neural networks, the implicit differentiation is employed to overcome computational barriers in backpropagation. The gradient-based approach for metamaterial design optimization is benchmarked against evolutionary strategies to explore global optima despite efficiency trade-offs. This co-optimization process represents a form of robotic evolution that enhances environmental sensing capabilities.
Our approach addresses these challenges through the development of physical intelligence (PI), which functions as an intelligent body interacting with and adapting to environmental stimuli, complementing computational intelligence (CI) which serves as the brain. The proposed system employs mechanical metamaterials with adaptable properties, utilizing a single central accelerometer as the sole electronic sensor. When external forces are applied to the metamaterial platform, mechanical wave propagation is detected and analyzed. Neural networks are trained to classify interaction sources or estimate properties of external objects based on these measurements.
Concurrently, the metamaterial structure undergoes optimization to enhance neural network performance through a two-tier framework: design optimization (high-level) and neural network training (low-level). The metamaterials are well known as their programmability and adaptability. Previous research proposed the effectiveness of inverse design of metamaterials to achieve target properties (e.g. maximizing energy absorption) using differentiable simulations of metamaterials. Inspired by the efficiency of differentiable simulation to optimize the designs, we implement differentiable simulation of metamaterials to obtain gradient information with respect to design parameters. Since the high-level optimization for design includes the iterative computational graph of training neural networks, the implicit differentiation is employed to overcome computational barriers in backpropagation. The gradient-based approach for metamaterial design optimization is benchmarked against evolutionary strategies to explore global optima despite efficiency trade-offs. This co-optimization process represents a form of robotic evolution that enhances environmental sensing capabilities.
–
Presenters
-
Bolei Deng
- Georgia Institute of Technology