Intuitive Engineering as a Proxy Facilitates the Evolutionary Optimization of an Underactuated Robot Ball

Giacomo Spigler, Simone Silenzi, Bart Horstman, Heike Vallery

Keywords: deep reinforcement learning, evolutionary robotics, intuitive engineering, underactuated robotics, virtual model control, mujoco, evolutionary optimization

TL;DR: A multi-stage optimization of the mechanical design and control of an underactuated robotic ball, "Fizzy", using evolutionary algorithms, intuitive nonlinear control, and deep reinforcement learning.

Abstract:

Underactuation can enable low-cost, light-weight robotics. However, their design is challenging. While classical engineering intuition often leads to reasonable hardware and control choices that ensure basic functionality, the resulting performance is usually low. In contrast, purely data-driven co-evolution of hardware and software conventionally needs high computational effort to deliver meaningful results.
We propose to leverage the advantages of both approaches by using classical intuitive controllers as proxies.
As an example, we consider ``Fizzy,'' an underactuated robotic ball that leverages a unique single-motor configuration in combination with dynamic imbalance for movement.
In a first optimization, an intuitive Virtual Model Control (VMC) proxy serves to quickly evaluate various design parameters like motor mass and axle positioning for a Covariance Matrix Adaptation Evolution Strategy (CMA-ES). The optimized configurations then serve as a foundation for training more sophisticated deep reinforcement learning (DRL) controllers.
Our methodology underscores the potential of integrating intuitive proxies with evolutionary algorithms to enhance the performance and efficiency of underactuated robotic systems, paving the way for more adaptable and cost-effective robotic designs.

Submission Number: 9