Abstract: Learning from both observations and self-exploration is a key capability of intelligent agents. It is present to various degrees in a number of animals that can combine information distilled from observing a task being performed with further knowledge acquired through self-exploration that contains variations over the demonstrations. We explore this mechanism in robotic arms in the context of learning to execute multi-step manipulation tasks. Specifically, the system accumulates geometric information on how humans typically manipulate objects with simple geometric shapes through learning from demonstrations. This geometric knowledge base is leveraged by a planner to find manipulation strategies given a novel scene. In addition, the new environment presents uncertainty in the values of the physical properties of the objects, such as their mass. We present an agent that has access to a simulator of the environment, instantiated by a physics engine, and uses it to explore multiple manipulation strategies and find successful sequences using Monte Carlo Tree Search. The search runs multiple forward simulations of the novel scene by hypothesizing plausible values for the unknown parameters of the mass of the objects. This abstract presents a preliminary implementation and results of this system, in which a dual-arm 16-DOF manipulator finds novel manipulation strategies by combining a geometric knowledge base learned from demonstrations with self-exploration in hypothesized simulated worlds in a simple task.

Cite as: Claudia Pérez-D'Arpino, Julie A. Shah. Observational and self-learning of multi-step robotic manipulation with unknown physical properties. Workshop on Modeling the Physical World: Perception, Learning, and Control. 32nd Conference on Neural Information Processing Systems (NeurIPS 2018). 2018.