A comparative study of reinforcement learning mechanisms for robot online adaptation under progressive damage

This thesis investigates the effectiveness of reinforcement learning mechanisms for online behavioral adaptation in autonomous mobile robots subjected to progressive sensor degradation. Using the ARGoS robotic simulator and a footbot equipped with proximity sensors and differential drive actuators, we examine how four exploration-exploitation strategies enable a neural network controller to maintain an obstacle avoidance behaviour under varying degrees of sensory failure. Controllers are initialized through the use of offline evolutionary optimization via a genetic algorithm (GALib), providing a population of high-performing recurrent neural network configurations as a starting repertoire for subsequent online adaptation. Experiments are then conducted in an online setting across seven damage scenarios, progressively disabling between one and seven of the robot’s proximity sensors. Performance is evaluated along two primary axes: a fitness function combining forward velocity with obstacle proximity, and collision counts across adaptation phases. Overall, the findings highlight both the potential and the limitations of lightweight online adaptation mechanisms in dynamic and fault-prone robotic environments.