Creative Problem Solving by Robots Using Action Primitive Discovery

Humans and many other species have the remarkable ability to innovate and creatively problem solve on-the-fly. Inspired by these abilities, we propose a framework for action discovery in problem solving scenarios similar to puzzle-boxes used to evaluate intelligence in animal species. The proposed framework assumes that the robot starts with a knowledge base including predicates and actions, which, however, are insufficient to solve the problem faced by the robot. We describe a method for discovering new action primitives through object exploration and action segmentation, which is able to iteratively update the robot's knowledge base on-the-fly until the solution becomes feasible. We implemented and evaluated the framework using a 3D physics-based simulated object retrieval task for the Baxter bi-manual robot. Results suggest that action segmentation is one viable path towards enabling autonomous agents to adapt on-the-fly and in short amounts of time to new situations that were unforeseen by their programmers and engineers.