NeuroMesh: A Unified Neural Inference Framework for Decentralized Multi-Robot Collaboration

A research team of 11 authors led by Yang Zhou has developed NeuroMesh, a groundbreaking framework that addresses the persistent challenge of deploying learned AI models across heterogeneous robot teams. The system creates a unified pipeline that standardizes observation encoding, message passing, aggregation, and task decoding—critical components that previously varied widely between different robot platforms. NeuroMesh's dual-aggregation paradigm enables both reduction- and broadcast-based information fusion, while its parallelized architecture decouples cycle time from end-to-end latency, allowing for more efficient real-time collaboration.

The framework's high-performance C++ implementation leverages Zenoh for robust inter-robot communication and supports hybrid GPU/CPU inference, making it adaptable to various hardware configurations. The researchers validated NeuroMesh on mixed teams of aerial and ground robots across three key applications: collaborative perception, decentralized control, and task assignment. The system demonstrated robust operation across diverse task structures and payload sizes, proving its versatility for real-world deployment scenarios. With acceptance at IEEE Robotics Automation Letter (RA-L) and plans for open-source release, NeuroMesh represents a significant step toward standardized AI deployment in multi-robot systems.