In this paper we present a neural architecture to learn a bi-directional mapping between actions and language. We implement a Multiple Timescale Long Short-Term Memory (MT-LSTM) network comprised of 7 layers with different timescale factors, to connect actions to language without explicitly learning an intermediate representation. Instead, the model self-organizes such representations at the level of a slow-varying latent layer, linking action branch and language branch at the center. We train the model in a bi-directional way, learning how to produce a sentence from a certain action sequence input and, simultaneously, how to generate an action sequence given a sentence as input. Furthermore we show this model preserves some of the generalization behaviour of Multiple Timescale Recurrent Neural Networks (MTRNN) in generating sentences and actions that were not explicitly trained. We compare this model with a number of different baseline models, confirming the importance of both the bi-directional training and the multiple timescales architecture. Finally, the network was evaluated on motor actions performed by an iCub robot and their corresponding letter-based description. The results of these experiments are presented at the end of the paper.

This paper addresses the problem of having a robot executing motor tasks requested by a human through spoken language. Verbal instructions do not typically have a one-to-one mapping to robot actions, due to various reasons: economy of spoken language, e.g., one short instruction might indeed correspond to a complex sequence of robot actions, and details about action execution might be omitted; grounding, e.g., some actions might need to be added or adapted due to environmental contingencies; embodiment, e.g., a robot might have different means than the human ones to obtain the goals that the instruction refers to. We propose a general cognitive architecture to deal with these issues, based on three steps: i) language-based semantic reasoning on the instruction (high-level), ii) formulation of goals in robot symbols and probabilistic planning to achieve them (mid-level), iii) action execution (low-level). The description of the mid-level is the main focus of this paper. The robot plans are adapted to the current scenario, perceived in real-time and continuously updated, taking in consideration the robot capabilities, modeled through the concept of affordances: this allows for flexibility and creativity in the task execution. We showcase the performance of the proposed architecture with real world experiments using the iCub humanoid robot, also in the presence of unexpected events and action failures.