Software Product Lines for Development of Evolutionary Robots

Evolutionary Robotics utilizes evolutionary algorithms for training robot controllers (e.g., neural networks) and adapting robot morphologies for different environments in design and runtime. One of the main challenges in robotics is the lack of reusability as AI-based robot controllers have to be trained from scratch for any change in the environment or a new task specification that a robot should adapt to. Training Artificial Neural Networks can be computationally heavy, time-consuming, and hard to reuse due to their monolithic black-box nature. The building blocks of emerging behaviors from Artificial Neural Networks cannot be fully separated or reused. We address the issue of reusability and propose an incremental approach for applying the reusability of behaviors. We implemented an Evolutionary Robotics framework to form a product family of robots. This product family is used to show the feasibility of our method for handling variability in a domain. Our results can be used to demonstrate a sample binding between the software product lines and machine learning domains.