TANGO: Commonsense Generalization in Predicting Tool Interactions for Mobile Manipulators

Robots assisting us in factories or homes must learn to make use of objects\nas tools to perform tasks, e.g., a tray for carrying objects. We consider the\nproblem of learning commonsense knowledge of when a tool may be useful and how\nits use may be composed with other tools to accomplish a high-level task\ninstructed by a human. We introduce a novel neural model, termed TANGO, for\npredicting task-specific tool interactions, trained using demonstrations from\nhuman teachers instructing a virtual robot. TANGO encodes the world state,\ncomprising objects and symbolic relationships between them, using a graph\nneural network. The model learns to attend over the scene using knowledge of\nthe goal and the action history, finally decoding the symbolic action to\nexecute. Crucially, we address generalization to unseen environments where some\nknown tools are missing, but alternative unseen tools are present. We show that\nby augmenting the representation of the environment with pre-trained embeddings\nderived from a knowledge-base, the model can generalize effectively to novel\nenvironments. Experimental results show a 60.5-78.9% absolute improvement over\nthe baseline in predicting successful symbolic plans in unseen settings for a\nsimulated mobile manipulator.\n