Decentralized Swarm learning informed by Spectral Collective Perception
Keywords: Swarm Robotics, Decentralized Spectral Analysis, Neuroevolution, Cooperative Object Transport, Social Learning, Swarm Intelligence, Quality-Diversity algorithms
Abstract:
Swarm robotics [1] investigates decentralized, collaborative behaviors in large groups of simple robots inspired by natural systems like ants or bees [2,3]. In this field, the diffusion of local information among agents through communication is critical for enabling collective decision-making and adaptation to tasks and environments [4]. Recently, we proposed a method termed Spectral Swarm Robotics (SSR) [5] allowing robotic swarms to decentrally estimate the global shape of their arena through strictly local interactions, opening the way towards a rich collective perception toolkit. This approach relied on local robot communication to perform decentralized spectral shape analysis, a paradigm traditionally only computable in a centralized manner [6,7].
We propose to extend this approach so that 1) the swarm can identify in a decentralized way the geometries of all objects in a given arena (e.g. cylinder, square, star, arbitrarily-shaped objects). In turn, 2) the shape-invariant spectral fingerprints of each object will be used as direct inputs to drive collective control and decision-making (e.g., selecting and transporting specific objects to goals). The combination of these two objectives will allow swarms where global geometrical knowledge becomes accessible on individual robots to inform control.
These objectives will be realised by adapting the original SSR algorithm to local communication subgraphs around objects, with dedicated motility strategies to aggregate around objects. These fingerprints are then diffused through the swarm via local consensus and consumed by decentralized controllers: artificial neural networks on each robot. The latter will be trained to solve a cooperative manipulation task [13,14] (detect and push specific objects to target sites), using either hand-crafted algorithms or training methods. Cf proposed workplan below.
Experimental validation of developed algorithms will be achieved using Pogobot robots [11] ( https://pogobot.github.io/ ), a novel open-hardware swarm robotics platform developed at the ISIR lab. This framework is complemented by a dedicated simulator (Pogosim [12] at https://github.com/Adacoma/pogosim ) allowing for fast prototyping (exact same code used on both simulations and real robots).
Our findings aim to establish the emerging field of spectral swarm robotics as a foundation for tackling advanced applications like collective sensing, decision-making, and decentralized manipulation. This work represents a transformative step toward extending swarm robotics methods to complex tasks and real-world scenarios.
Proposed workplan:
Phase 1 — Object identification by aggregation (Months 1–2)
Goal: form a local sub-swarm (halo) around each physical object and label it. Build a decentralized system that detects and clusters robots around arbitrarily shaped objects in an arena and assigns each object a local “sub-swarm.” For each object, use SSR (decentralized spectral sensing with moving robots) to compute a compact fingerprint and diffuse this knowledge through the swarm.
- Aggregation around objects:
- V1 — Active-walls (fastest prototype): treat objects as “active walls” that periodically broadcast a simple “I am a wall” message. Robots aggregate around the emitter boundary.
- V2 — Passive sensing (no broadcast): use either (a) a magnetometer (for absolute heading stability) or (b) stationary light levels from photosensors; confirm by persistent contact/drag.
- Stop vs. move: implement aggregation+stop (simplest) and aggregation+move (robots continue patrolling while maintaining a boundary halo).
- Sub-swarm specialization: run the SSR [5] algorithm locally around each object and set a distinct LED color (w.r.t. object geometry) or comms channel per object to avoid cross-contamination when objects are close. Tune timings so halo formation and labeling are fast and stable.
Milestones: reliable boundary detection; stable halo formation around multiple nearby objects; per-object color/channel assignment; short demo video.
Phase 2 — Decentralized pushing to a precise direction/goal (Months 3–4)
Goal: move a chosen object toward a designated zone (e.g., a bright corner) with minimal sensing [13,14].
- Goal detection & diffusion: a fraction of robots disperse to find the target zone via a light gradient; they then diffuse this information (scalar “goal potential”) to the whole swarm using standard consensus/diffusion.
- Simple pushing controller with role separation:
- The swarm agrees on the push direction: consensus on the goal-gradient measured in the object’s halo.
- Rear robots align with the push vector and apply force; side robots align tangentially to steer and prevent drift; front robots yield/clear the path.
- Insert collective pause/re-align events: robots stop together on a coarse timer or on obstacle detection, re-estimate gradients, then resume.
- Learning-based controller: train a frugal MLP policy (≤ a few hundred parameters) that maps local features to actions (collaborative object transport); and validate experimentally.
- Inputs: SSR collective sensing like per-object spectral fingerprint and goal potential, communication with neighbors, angle to objects, etc
- Training method: initial neuroevolution offline training in simulation [8,10], then optional decentralized online-training [9] in experiments to adapt policies to real world settings (sim2real adaptation)
Milestones: repeated trials showing successful transport to the target, with logs of time-to-goal, stops/re-align counts, and interference with non-target objects; trained MLP policy; short demo.
Phase 3 — Writing & packaging (Months 5-6)
Final report/manuscript writing. If time allows: conference/journal article submission.
Internship terms and conditions:
This 5/6-month interdisciplinary M2 internship is at a crossroad between swarm robotics, artificial intelligence, and mathematical models of information diffusion. It will take place at the ISIR lab (Institut des Systèmes Intelligents et de Robotique) at Sorbonne Université, among an interdisciplinary team composed of computer scientists, roboticists and biophysicists. It will be funded by the “Spectral-Swarm-Robotics” ANR project.
It will be jointly supervised by Dr Leo Cazenille and Pr Nicolas Bredeche (ISIR, SU).
Potential publication opportunities in AI/robotics conferences.
Expected skills and research interests:
- Master 2 level in AI / Computer Science, Robotics, Bio-physics or Applied Mathematics
- Proficiency in C and Python
- Optional knowledge of data-science (pandas, numpy, etc), plotting (matplotlib/seaborn), machine learning libraries, such as pytorch, etc..
References (most important in bold):
[1] H Hamann. Swarm robotics: A formal approach, volume 221. 2018.
[2] M Dorigo, G Theraulaz, and V Trianni. Reflections on the future of swarm robotics. Science Robotics, 2020.
[3] M Rubenstein, A Cornejo, and R Nagpal. Programmable self-assembly in a thousand-robot swarm. Science, 2014.
[4] Cazenille, L., Toquebiau, M., Lobato-Dauzier, N., Loi, A., Macabre, L., Aubert-Kato, N., ... & Bredeche, N. (2025). Signalling and social learning in swarms of robots. Philosophical Transactions A, 383(2289), 20240148.
[5] Cazenille, L, N Lobato-Dauzier, A Loi, M Ito, O Marchal, Aubert-Kato, N, Bredeche, N, and Genot, AJ. Hearing the shape of an arena with spectral swarm robotics. arXiv:2403.17147, 2024.
[6] D Spielman. Spectral graph theory. Combinatorial scientific computing, 2012.
[7] Reuter M., Wolter F.-E., Peinecke N., “Laplace–Beltrami spectra as Shape-DNA”
[8] Hasselmann, K., Ligot, A., Ruddick, J., & Birattari, M. (2021). Empirical assessment and comparison of neuro-evolutionary methods for the automatic off-line design of robot swarms. Nature communications, 12(1), 4345.
[9] Bredeche, N., & Fontbonne, N. (2022). Social learning in swarm robotics. Philosophical Transactions of the Royal Society B, 377(1843), 20200309.
[10] Chatzilygeroudis, Konstantinos, et al. "Quality-diversity optimization: a novel branch of stochastic optimization." Black Box Optimization, Machine Learning, and No-Free Lunch Theorems. Cham: Springer International Publishing, 2021. 109-135.
[11] Loi, A., Macabre, L., Fersula, J., Amini, K., Cazenille, L., Caura, F., Guerre, A., Gourichon, S., Dauchot, O. and Bredeche, N., 2025. Pobogot--An Open-Hardware Open-Source Low Cost Robot for Swarm Robotics. arXiv preprint arXiv:2504.08686.
[12] Cazenille, L., Macabre, L. and Bredeche, N., 2025. Pogosim-a Simulator for Pogobot robots. arXiv preprint arXiv:2509.10968.
[13] Groß, R. and Dorigo, M., 2009. Towards group transport by swarms of robots. International Journal of Bio-Inspired Computation, 1(1-2), pp.1-13.
[14] Tuci, Elio, Muhanad HM Alkilabi, and Otar Akanyeti. "Cooperative object transport in multi-robot systems: A review of the state-of-the-art." Frontiers in Robotics and AI 5 (2018): 59.