We consider the problem of robots planning for complex, long-horizon tasks. Several algorithms exist to solve arbitrary planning problems, but their running time scales poorly with the plan length. However, knowledge learned from experience can be used to guide planning to more quickly find a sequence of actions leading to the goal. We generalize the convolutional neural network to apply to graphstructured input and train it to learn the cost of reaching a goal from a given state encoded as a graph. This heuristic can be integrated with a conventional planning algorithm to pick intermediate goals that are likely to eventually lead to the final goal. As a result, the planning algorithm never has to plan all the way to the goal, leading to faster plan generation. We implement and evaluate this approach in a simulated 3D environment and eventually on a physical robot.

“I am participating in SuperUROP because I want to continue to build on my prior research in robotics and learning and eventually determine a path forward for graduate school and beyond.”