Search

This thesis proposes a novel and robust online adaptation mechanism for threat prediction and prevention capable of taking into consideration complex contextual and temporal information in its internal learning processes. The proposed mechanism is a hybrid cognitive computational model named SAFEL (Situation-Aware FEar Learning), which integrates machine learning algorithms with concepts of situation-awareness from expert systems to simulate both the cued and contextual fear-conditioning phenomena. SAFEL is inspired by well-known neuroscience findings on the brain's mechanisms of fear learning and memory to provide autonomous robots with the ability to predict undesirable or threatening situations to themselves. SAFEL's ultimate goal is to allow autonomous robots to perceive intricate elements and relationships in their environment, learn with experience through autonomous environmental exploration, and adapt at execution time to environmental changes and threats. SAFEL consists of a hybrid architecture composed of three modules, each based on a different approach and inspired by a different region (or function) of the brain involved in fear learning. These modules are: the Amygdala Module (AM), the Hippocampus Module (HM) and the Working Memory Module (WMM). The AM learns and detects environmental threats while the HM makes sense of the robot's context. The WMM is responsible for combining and associating the two types of information processed by the AM and HM. More specifically, the AM simulates the cued conditioning phenomenon by creating associations between co-occurring aversive and neutral environmental stimuli. The AM represents the kernel of emotional appraisal and threat detection in SAFEL's architecture. The HM, in turn, handles environmental information at a higher level of abstraction and complexity than the AM, which depicts the robot's situation as a whole. The information managed by the HM embeds in a unified representation the temporal interactions of multiple stimuli in the environment. Finally, the WMM simulates the contextual conditioning phenomenon by creating associations between the contextual memory formed in the HM and the emotional memory formed in the AM, thus giving emotional meaning to the contextual information acquired in past experiences. Ultimately, any previously experienced pattern of contextual information triggers the retrieval of that stored contextual memory and its emotional meaning from the WMM, warning the robot that an undesirable situation is likely to happen in the near future. The main contribution of this work as compared to the state of the art is a domain-independent mechanism for online learning and adaptation that combines a fear-learning model with the concept of temporal context and is focused on real-world applications for autonomous robotics. SAFEL successfully integrates a symbolic rule-based paradigm for situation management with machine learning algorithms for memorizing and predicting environmental threats to the robot based on complex temporal context. SAFEL has been evaluated in several experiments, which analysed the performance of each module separately. Ultimately, we conducted a comprehensive case study in the robot soccer scenario to evaluate the collective work of all modules as a whole. This case study also analyses to which extent the emotional feedback of SAFEL can improve the intelligent behaviour of a robot in a practical real-world situation, where adaptive skills and fast/flexible decision-making are crucial.