Onni Wuoti

Autonomous robots often work under unstable and changing conditions where decision making and navigation tasks increase in complexity due to the uncertainties, and require evaluation of multiple information sources at once. In reinforcement learning information about the environment is often compressed to a single scalar reward, which makes evaluating different components not possible. This is a valuable property to have in tasks with uncertain environments such as search and rescue operations, environmental monitoring, or general exploration of unconventional spaces.  This research aims to create a Multi-Objective Reinforcement Learning (MORL) framework based on biomimicry, where we copy the idea of animals weighing the importance of different signals before making the decision. In this case this would be done by storing rewards in a vector, where components represent different rewards. This way, different components could be weighed independently, and context sensitive decision making can be done.  By providing these different reward channels in the form of vectors the information of the stakeholder can be increased, which in turn creates deeper understanding and trust within the robots decision-making. This also makes evaluation and further development more feasible, as more information is available. This framework is then implemented into a multi-robot system, where global and local knowledge of the robots will be combined to create a shared knowledge of their surrounding environment.