Abstract
Enabling machines to respond appropriately to natural language commands couldgreatly expand the number of people to whom they could be of service. Recently,advances in neural network-trained word embeddings have empowered non-embodiedtext-processing algorithms, and suggest they could be of similar utility forembodied machines. Here we introduce a method that does so by training robotsto act similarly to semantically-similar word2vec encoded commands. We showthat this enables them to act appropriately, after training, topreviously-unheard commands. Finally, we show that inducing such an alignmentbetween motoric and linguistic similarities can be facilitated or hindered bythe mechanical structure of the robot. This points to future, large scalemethods that find and exploit relationships between action, language, and robotstructure.