ReIL: A Framework for Reinforced Intervention-based Imitation Learning

Compared to traditional imitation learning methods such as DAgger and DART, intervention-based imitation offers a more convenient and sample efficient data collection process to users. In this paper, we introduce Reinforced Intervention-based Learning (ReIL), a framework consisting of a general intervention-based learning algorithm and a multi-task imitation learning model aimed at enabling non-expert users to train agents in real environments with little supervision or fine tuning. ReIL achieves this with an algorithm that combines the advantages of imitation learning and reinforcement learning and a model capable of concurrently processing demonstrations, past experience, and current observations. Experimental results from real world mobile robot navigation challenges indicate that ReIL learns rapidly from sparse supervisor corrections without suffering deterioration in performance that is characteristic of supervised learning-based methods such as HG-Dagger and IWR. The results also demonstrate that in contrast to other intervention-based methods such as IARL and EGPO, ReIL can utilize an arbitrary reward function for training without any additional heuristics.