核心概念
This paper presents a novel method for the co-design of soft robot morphology and control using a CUDA-accelerated neural evolution approach integrated with Large Language Model (LLM) supervision to significantly improve the efficiency and diversity of the evolutionary process.
摘要
The paper addresses the challenge of co-designing the morphology and control of soft robots. The authors propose an innovative method that dual-encodes the morphological and control designs within a multilayer perceptron (MLP) and utilizes a neural evolution approach to generate new MLP configurations.
To address the limitations of traditional evolutionary algorithms, the authors leverage the capabilities of the GPT-4-Turbo Large Language Model to guide the evolution process. The LLM supervision helps maintain a high level of diversity in the evolved robot morphologies and control strategies, while also accelerating the rate of fitness convergence.
The authors employ a 3D voxel-based soft robot model that allows for various design possibilities, including empty spaces, expanding and contracting muscles, and support structures made from soft tissue or hard bone material. To enhance the neural network's understanding of the robot's morphology, the authors introduce Gaussian positional encoding, which maps the spatial query input into higher dimensions.
The CUDA-accelerated neural evolution framework enables the authors to perform up to 7,892,537,853 spring-mass physics simulations per second within a single NVIDIA RTX-3090 GPU, significantly speeding up the co-design process.
The results demonstrate the effectiveness of the proposed approach, showing that the LLM supervision maintains a high level of diversity in the evolved robot morphologies and control strategies, while also accelerating the rate of fitness convergence compared to the traditional evolutionary algorithm.
統計資料
The paper provides the following key metrics and figures:
Up to 7,892,537,853 spring-mass physics simulations per second within a single NVIDIA RTX-3090 GPU
General parameters:
Generations: 100
Population Size: 30
Robot Size: 5 x 5 x 5
Repetitions: 3
Initial Mutation Rate: 0.1
Initial Mutation Scale: 0.1
Initial Crossover Rate: 0.4
Initial Elite Percentage: 0.3
Simulation parameters:
Gravity: 9.81 m/s^2
Mass per Indices: 0.1 Kg
Original Rest Length: 0.1 m
Simulation Time-Step: 10^-5 s
Muscle Spring Constant: 2 x 10^3 N/m
Soft Tissue Spring Constant: 10^3 N/m
Hard Bone Spring Constant: 10^4 N/m
Spring Damping Ratio: 0.1
Muscle Max Amplitude: 0.25
Muscle Max Phase: π
Plane Elastic Coefficient: 10^5 N/m
Plane Damping Ratio: 0.1
Plane Static Coefficient of Friction: 0.6
Plane Kinetic Coefficient of Friction: 1.0