Generalization and Transfer Learning in Noise-Affected Robot Navigation Tasks

When a robot learns to solve a goal-directed navigation task with reinforcement learning, the acquired strategy can usually exclusively be applied to the task that has been learned. Knowledge transfer to other tasks and environments is a great challenge, and the transfer learning ability crucially depends on the chosen state space representation. This work shows how an agent-centered qualitative spatial representation can be used for generalization and knowledge transfer in a simulated robot navigation scenario. Learned strategies using this representation are very robust to environmental noise and imprecise world knowledge and can easily be applied to new scenarios, offering a good foundation for further learning tasks and application of the learned policy in different contexts.