Learning Constraints in Robot Trajectories

Constraint learning is a research area covering different methodologies for an inductive construction of constraint theories from data. In many contexts, a system can benefit from constraint learning to optimize the process in terms of time and performance. In this thesis, trajectories data of a trained reinforcement learning agent are used as examples to define constraints characterizing the environment. The methodology involves three main steps. In the first phase an agent is trained in a two-dimensional space containing unsafe areas. Then, the agent is tested and its trajectories are stored and used as data for learning a formula that classifies the region of the training examples as positive, thereby identifying safe and unsafe areas. In the last step a new agent is trained by using the formula to assign a penalty to movements towards an unsafe area. The addition of prior knowledge about the environment leads to a significant improvement of the agent's performance in terms of satisfaction of safety specifications. Hence, in fields like autonomous robotics, where safety concerns are paramount, systems can benefit from constraint learning to increase the level of reliability and guarantee a "safer" exploration.