Abstract
Effective coordination in multi-agent systems, such as robotic search-and-rescue teams, requires sharing spatial information under significant communication constraints. We propose a three-level framework grounded in predictive coding that enables decentralized agents to learn a sparse, utility-driven communication protocol. Our results demonstrate that agents spontaneously develop neural representations analogous to biological grid cells and social place cells. The system maintains high navigation success rates even when bandwidth is reduced to 4 bits per step, achieving a 32:1 compression ratio of spatial maps.
Introduction
In partially observable environments, agents must communicate to resolve uncertainty and coordinate exploration. However, broadcasting raw sensory data is often infeasible due to bandwidth limitations. Existing methods frequently fail to prioritize information based on its semantic value to the receiver. This paper addresses the challenge of learning a decentralized, intelligent communication mechanism that transmits only essential data to resolve a partner's uncertainty.
Methodology
The proposed framework utilizes predictive coding—the principle that agents generate internal models of the world and update them based on prediction errors—across three hierarchical levels.
1. Individual Predictive Model
Each agent employs a Transformer-based architecture to map egocentric RGB images into a Bird's-Eye-View (BEV) representation. An LSTM-based path integrator predicts the agent's future pose from its velocity. This self-supervised motion prediction task encourages the network to develop a stable spatial metric.
2. Social Predictive Coding and Communication
Agents use a dual-stream LSTM to predict both their own and their partner's future states. Communication is governed by a Variational Information Bottleneck (VIB), which acts as a mathematical filter. The VIB forces the sender to compress its spatial map into a minimal latent message that maximizes the receiver's ability to predict its own future state, effectively filtering for utility.
3. Strategic Exploration
Navigation is managed via Hierarchical Reinforcement Learning with an Intrinsic Curiosity Module (HRL-ICM). A high-level meta-controller selects goals that reduce collective team uncertainty, while a low-level A* planner executes the pathfinding.
Results and Discussion
Emergence of Biological Representations
A significant finding is the spontaneous emergence of brain-like structures without explicit biological hardcoding. Optimizing for motion prediction led to the development of artificial grid cells—neurons with hexagonal firing patterns used for spatial navigation. Furthermore, the requirement to predict a partner's state resulted in social place cells, specialized neurons that fire based on the partner's location, mirroring findings in mammalian hippocampi.
Communication Efficiency and Resilience
The framework demonstrated remarkable resilience to bandwidth constraints on the Memory-Maze benchmark. When bandwidth was restricted from 128 bits to 4 bits per step, the success rate only decreased from 73.5% to 64.4%, whereas a full broadcast baseline collapsed to 28.6%. The VIB mechanism achieved a 32:1 compression ratio while keeping reconstruction error below 5%.
Strategic Communication
Agents learned to communicate strategically, often broadcasting messages from dead-ends. These signals serve as high-utility "prediction errors" for partners, preventing redundant exploration. The severe bottleneck also forced the emergence of a discrete, symbolic language representing high-level contexts like "Dead-end" or "Target Found."
Conclusion
This work demonstrates that predictive coding is a powerful objective for learning decentralized multi-agent coordination. By focusing on uncertainty reduction, agents develop efficient communication and sophisticated spatial representations. Future work will explore continuous communication structures and deployment on physical robotic platforms to address real-world sensory noise.