Improving Robot Skills by Integrating Task and Motion Planning with Learning from Demonstration

Abstract

Collecting human demonstrations to train robots to solve real-world tasks is cumbersome. Having the robot generate its own demonstrations would enable autonomous data collection and obviate challenges associated with the embodiment gap. We present a novel approach to generalizing robot manipulation skills by combining a sampling-based task-and-motion planner with a learning-from-demonstration algorithm. Starting with a small library of scripted primitive skills (e.g. Push) and object-centric symbolic predicates (e.g. On(block, plate)), the planner autonomously generates a demonstration dataset of manipulation skills in the context of a long-horizon task. An offline policy learning algorithm then extracts a skill from the dataset without further interactions with the environment and replaces the scripted skill in the existing library. Refining the skill library improves the performance of the planner, which in turn facilitates data collection for more complex manipulation skills. We evaluate our approach in simulation, on three mobile manipulation tasks. We show that the proposed method is robust to suboptimal demonstrations and is suitable for long-horizon tasks. Furthermore, interaction with the environment is collision-free because of the use of planner demonstrations, making the approach more amenable to persistent robot learning in the real world.