Research Summary.

Multi-agent communication literature has primarily focused on learning symbolic communication through discrete communication channels (discrete action spaces). However, embodied agents have another expressive modality for communicating, through the use of body movement. Indeed this takes inspiration from human agents, who employ language (symbols) and gesture (movement) concurrently as co-expressive forms of communication. Within this thread of research, I am currently leading a project which explores how to enable embodied agents (such as the ones shown below) to learn to communicate through physical action (articulated motion). It is loosely motivated by problem domains that use robots or virtual characters. Two novel and interesting aspects of embodied communication are the use of high-dimensional continuous communication channels (high-dimensional continuous action spaces) and a non-cheap talk problem setting. The former, combined with self-play, implies a highly non-convex optimization landscape. The latter means there is a cost expenditure associated with generating physical communication. We exploit this cost for generating communication through physical movement — to regularize protocol learning and induce bias in the search for an optimal communication protocol.

The research questions we have begun to explore in this work are:

  • Under what minimal set of realistic common-knowledge constraints can we obtain zero-shot emergent communication?

  • How do we enable agents with an algorithm for automatically inferring optimal zero-shot communication protocols?

  • What are the implications as we scale to continuous domains and aim to induce zero-shot communication through an agent’s embodiment (communication through physical action)?

Zero-shot communication is the problem setting where the task for agents is to communicate effectively with novel partners (i.e. strangers), at test time.

Below are examples of the types of “embodied agents” being investigated. It is fundamentally motivated by the idea that embodied agents (such as humans and robots) implicitly communicate through movement. Yet most work in multi-agent communication focuses on symbolic (explicit) communication (e.g. through words). If we are to move towards human-level intelligence (which is an embodied intelligence), artificial embodied agents must be able to also communicate through this important channel. Particularly artificial agents intended to be compatible with humans or exist in more realistic human settings. Note: Robot arm agent is the only one used to date.

Robot Arm Agent

Robot Arm Agent

Humanoid Agent

Humanoid Agent

Relevant Publications and Preprints.

  • Quasi-Equivalence Discovery for Zero-Shot Emergent Communication (arXiv Preprint — Full Paper, 2021) [arXiv]

  • Exploring Zero-Shot Emergent Communication in Embodied Multi-Agent Populations (arXiv Preprint — Full Paper, 2020) [arXiv]

  • Towards Emerging Nonverbal Communication Protocols for Multi-Robot Populations (RSS Workshop — Extended Abstract, 2020) [PDF]

Relevant Talks.

  • Exploring Zero-Shot Emergent Communication in Embodied Multi-Agent Populations (NeurIPS Deep RL Workshop, Dec 2020) [Poster Talk — 3 min] [Poster]

  • Learning through Interaction in Multi Agent Systems (UMass Amherst ML and Friends Lecture Series, Oct 2020) [Seminar Talk — 1 hour]

  • Learning to Communicate Nonverbally for Embodied Agent Populations (ICML WiML Workshop, Jul 2020) [Facebook Sponsor Talk — 15 min]

  • Emerging Nonverbal Protocols for Multi-Robot Populations (RSS Workshop on Emergent Behaviors in Human-Robot Systems, Jul 2020) [Spotlight Talk — 5 min]