Evolution-Based Shape and Behavior Co-Design of Virtual Agents.

Area of Science:

• Robotics
• Artificial Intelligence
• Evolutionary Computation

Background:

• Autonomous moving agents require sophisticated design for effective locomotion.
• Simultaneous optimization of agent morphology and behavior is a significant challenge.

Purpose of the Study:

• To introduce a novel co-design method for optimizing autonomous agents' shape and locomotion.
• To leverage deep reinforcement learning and evolutionary algorithms for agent design.
• To incorporate user-defined constraints into the co-design process.

Main Methods:

• A hybrid approach combining deep reinforcement learning and evolutionary algorithms.
• Physics-based simulation for evaluating agent locomotion.
• Policy transfer between generations to accelerate training.
• Crossover and mutation operations for generating new agent designs.

Main Results:

• Significant performance improvements in locomotion tasks, up to 150% with greater design flexibility.
• Demonstrated effectiveness even with limited design modifications (50% improvement with 10% changes).
• Evolved agents exhibited diverse shapes and adaptive behaviors.
• Method achieves substantial results within 30 minutes on a single GPU.

Conclusions:

• The proposed co-design method effectively optimizes both the form and function of autonomous agents.
• The approach offers a computationally efficient solution for complex agent design problems.
• User control and evolutionary principles can be integrated for tailored agent development.