Simultaneous Optimization of Discrete and Continuous Parameters Defining a Robot Morphology and Controller.

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Robot design and controller creation are complex, time-consuming processes.
• Current machine learning methods for robot design often limit the search space by restricting part types.
• A broader search space is needed for discovering globally optimal robot designs.

Purpose of the Study:

• To develop a novel, efficient method for exploring diverse robot designs.
• To reduce the labor-intensive nature of robot morphology and controller design.
• To improve robot performance through automated design optimization.

Main Methods:

• Combines proximal policy optimization (PPO) or soft actor-critic (SAC) for controllers.
• Utilizes the REINFORCE algorithm for optimizing numerical parameters of rigid parts.
• Introduces a new method for determining the number and layout of robot parts and joints.

Main Results:

• The proposed method efficiently searches a richer set of robot designs compared to existing approaches.
• Experiments in physical simulations demonstrate superior performance in walking and manipulation tasks.
• The integrated optimization strategy outperforms simpler combinations of existing methods.

Conclusions:

• The novel method effectively automates robot design by exploring a vast design space.
• This approach significantly enhances robot performance for various tasks.
• The findings pave the way for more efficient and capable robot development.