Mechanisms for online adaptation in robots: A comparative study

Robots deployed in dynamic environments must be able to adapt autonomously to changing conditions and perturbations. This thesis examines online adaptation strategies for minimally cognitive robotic agents, with a focus on their ability to achieve and sustain high performance. We explore a range of adaptive controllers, including architectures inspired by Braitenberg vehicles and Artificial Neural Network-based strategies, from simple feed-forward topologies to recurrent networks with internal memory, each tested in navigation with a collision avoidance task. Our experimental results compare the performance of various mechanisms and adaptation policies, highlighting the trade-offs between reactivity, memory, and robustness in different online adaptation settings.