Read the local world
An observation encoder turns each agent’s private input into a compact embedding.
ICRA 2022 · Cooperative MARL
Emergent discrete message communication.
A compact, interpretable language for agents that need to act together— learned through broadcast, listening, and self-attention.
Communication architecture
01
DomainCooperative multi-agent RL
MechanismBroadcast + listen
RepresentationDiscrete messages
VenueIEEE ICRA 2022
01 The problem
Useful signals, readable language.
Communication is central to cooperation, but many multi-agent systems exchange continuous vectors whose expressive power comes at the cost of interpretability. The agents may coordinate, yet neither a human designer nor another system can easily inspect what was said.
This work asks whether agents can learn a discrete protocol—closer to a small vocabulary than an opaque latent channel—while preserving the performance benefits of learned communication.
Q / 01
Can agents learn messages small enough to inspect, yet expressive enough to cooperate?
02 The architecture
Speak. Listen. Act.
Each agent encodes its local observation, produces a candidate discrete message, and learns how strongly to attend to messages from other agents. The weighted message aggregate augments the agent’s own representation before its policy selects an action.
Broadcast and listen
Message logits select a discrete token while attention weights decide which incoming messages matter.
Act with shared context
The aggregated message joins the local embedding so the action policy can respond to the group.
03 What it showed
Interpretability without surrendering coordination.
A constrained channel can still carry the information a cooperative team needs.
Comparable to continuous communication
The learned discrete protocol achieved performance comparable to continuous-message approaches.
Far fewer symbols
Agents coordinated with a much smaller vocabulary, making the channel easier to inspect and reason about.
A channel for people, too
The same discrete interface enables humans to send messages interactively to participating agents.
04 / Paper
Published at ICRA 2022
Learning Emergent Discrete Message Communication for Cooperative Reinforcement Learning
IEEE International Conference on Robotics and Automation · 2022