Multi-goal policy learning for robotic manipulation is challenging. Prior successes have used state-based representations of the objects or provided demonstration data to facilitate learning. In this paper, by hand-coding a high-level discrete representation of the domain, we show that policies to reach dozens of goals can be learned with a single network using Q-learning from pixels. The agent focuses learning on simpler, local policies which are sequenced together by planning in the abstract space. We compare our method against standard multi-goal RL baselines, as well as other methods that leverage the discrete representation, on a challenging block construction domain. We find that our method can build more than a hundred different block structures, and demonstrate forward transfer to structures with novel objects. Lastly, we deploy the policy learned in simulation on a real robot.

We present a novel approach to goal recognition based on a two-stage paradigm of graph construction and analysis. First, a graph structure called a Goal Graph is constructed to represent the observed actions, the state of the world, and the achieved goals as well as various connections between these nodes at consecutive time steps. Then, the Goal Graph is analysed at each time step to recognise those partially or fully achieved goals that are consistent with the actions observed so far. The Goal Graph analysis also reveals valid plans for the recognised goals or part of these goals. Our approach to goal recognition does not need a plan library. It does not suffer from the problems in the acquisition and hand-coding of large plan libraries, neither does it have the problems in searching the plan space of exponential size. We describe two algorithms for Goal Graph construction and analysis in this paradigm. These algorithms are both provably sound, polynomial-time, and polynomial-space. The number of goals recognised by our algorithms is usually very small after a sequence of observed actions has been processed. Thus the sequence of observed actions is well explained by the recognised goals with little ambiguity. We have evaluated these algorithms in the UNIX domain, in which excellent performance has been achieved in terms of accuracy, efficiency, and scalability.