Internship in “GNNs Topological Design for Compact Communication Neurons in Multi-Robot Systems”

In the context of scarce knowledge and computation constraints, neural networks have become a viable way to model system dynamics and control actions thanks to their approximation capabilities. Within the domain of multi-robot systems, graph neural networks (GNNs) and their variants emerged as data-driven models for systems employing communication strategies [Li et Al., 2020]. Given that a group of communicating agents forms a graph, GNNs offer effective models for distributed estimation and control, even in the absence of detailed knowledge regarding individual robot dynamics. However, a notable drawback of these models lies in the substantial volume of communication variables required to achieve accurate predictions, surpassing those of their model-based counterparts [Blumenkamp et al., 2022]. Existing literature primarily addresses this challenge through the optimization of graph topology [Yang et al., 2021], focusing on minimizing the number of communication links within the graph.

Historically, neural networks took inspiration from biological systems and neuron models regulating the flow of currents in the neuronal cell membrane. Recently, Lechner et Al., 2020 proposed a novel architecture that combines the expressivity ability of the Liquid-Time-Constant network, a continuous neural model, with a sparse pattern in the neural network architecture. In particular, inspired by the nervous systems of C. elegans nematodes, the architecture divides the neurons composing the network in sensory, command and motor neurons, sparsely connected to each others. Although serving different purposes in driving the sensor input to the predicted outputs, each neuron employ the same dynamics. This architecture has proven highly effective in controlling autonomous agents for diverse tasks and demonstrates remarkable generalization capabilities beyond the training distribution [Chahine et al., 2023].

Motivated by the success of this innovative topology, the primary objective of this internship is to develop a novel neural network architecture wherein only a limited number of neurons are necessitated to communicate encoded information effectively. An example of such topology is illustrated in Figure 1. A possible architecture can take inspiration from the communication mechanisms of living beings, including birds or fishes, which deploy dedicated mechanisms for communication and cooperation with other animals/agents [Herzing et Al., 2015, Bass et Al., 2003,Avena-Koenigsberger et Al., 2018]. The ultimate goal of this work is to develop a general machinery to learn autonomous policy to generate control actions and estimations tailored for the complex dynamics inherent in multi-robot systems. In this respect, the candidate will implement benchmarks to train and evaluate the proposed learning model. Subsequently, rigorous testing will be conducted on a team of UAVs within a designated flight arena, deploying one of the use cases. Through these iterative steps, we aim to not only validate the efficacy of the developed model but also to demonstrate its robustness in real-world scenarios, thereby advancing the frontier of autonomous robotics.

References
◦ Lechner, Mathias, et al. “Neural circuit policies enabling auditable autonomy.” Nature Machine Intelligence 2.10 (2020): 642-652.
◦ F. Yang and N. Matni, “Communication Topology Co-Design in Graph Recurrent Neural Network based Distributed Control,” 2021 60th IEEE Conference on Decision and Control (CDC), Austin, TX, USA, 2021, pp. 3619-3626, doi: 10.1109/CDC45484.2021.9683779.
◦ Avena-Koenigsberger, Andrea, Bratislav Misic, and Olaf Sporns. “Communication dynamics in complex brain networks.” Nature reviews neuroscience 19.1 (2018): 17-33.
◦ Bass, Andrew H., and Jessica R. McKibben. “Neural mechanisms and behaviors for acoustic communication in teleost fish.” Progress in neurobiology 69.1 (2003): 1-26.
◦ Herzing, Denise L. “Making sense of it all: multimodal dolphin communication.” (2015).
◦ Blumenkamp, Jan, et al. “A framework for real-world multi-robot systems running decentralized GNN-based policies.” 2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022.
◦ Li, Qingbiao, et al. “Graph neural networks for decentralized multi-robot path planning.” 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2020.
◦ Chahine, Makram, et al. “Robust flight navigation out of distribution with liquid neural networks.” Science Robotics 8.77 (2023): eadc8892.

Contact Antonio Marino at antonio.marino@irisa.fr, providing Curriculum Vitae
(CV) and transcript of record. All the candidates will go through an interview process.