Scalable Co-optimization of Collective Robotic Mobility and the Artificial Environment

Projektinformationen

gAIa

ID Finanzhilfevereinbarung: 949940

DOI

10.3030/949940

EK-Unterschriftsdatum 7 Oktober 2020

Startdatum 1 Januar 2021

Enddatum 30 September 2026

Finanziert unter

• EXCELLENT SCIENCE - European Research Council (ERC)

Gesamtkosten

€ 1 495 338,00

EU-Beitrag

€ 1 495 338,00

Koordiniert durch

THE CHANCELLOR MASTERS AND SCHOLARS OF THE UNIVERSITY OF CAMBRIDGE
United Kingdom

Periodic Reporting for period 3 - gAIa (Scalable Co-optimization of Collective Robotic Mobility and the Artificial Environment)

Berichtszeitraum: 2024-01-01 bis 2025-06-30

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

The premise of this project is that the environment is as much a variable as the robot itself. I want to expose the coupling between environmental structure and collective robotic mobility. In pursuit of this goal, I propose a co-optimization scheme that finds the best robot-environment pairs in an automated, scalable manner. The work in this project will (i) optimize control policies that define the behavior of collective mobile robot systems, and (ii) find layouts and structures of environments that are more conducive to efficient coordination and cooperation. The developed techniques will allow us to perform first-of-a-kind analyses that would reveal novel spatial topologies and morphologies, and the collective robot policies optimized around them. Ultimately, this project will spearhead new ways of thinking about transport planning and urban design, in the wake of a new generation of mobile vehicles that are connected and coordinated.

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

We have developed techniques that facilitate the learning of complex cooperative behaviors for multi-robot and multi-agent systems. We have also pioneered the first usages of Graph Neural Network-based policies in real-world robot deployments.

We have formalized the environment optimization problem for multi-agent navigation. Our first paper provides the conditions and proofs under which we can guarantee completeness of agent pathfinding. We show how environment optimization indeed provides significant benefits to the team, under a small cost. A follow-up journal paper, extends this theory to prioritized and constrained robots. The results are again backed up by theory that details the conditions under which priorities and path completeness will be met.

We have provided a formal problem statement for the co-optimization problem, and developed a hybrid learning-optimization approach to solving it. We demonstrate results on agent teams with limited communication, and show the value of joint policy-environment optimization. In our results, we show that counterintuitively, adding obstacles and clutter can guide autonomous agents in pursuit of more efficient and safer paths.

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

(1) Our work has pioneering the use of differentiable communication through GNNs, as applied to robot systems. This has lead to a break-through for deployment of robot networks, as we can now much more effectively and efficiently synthesize coordinated behaviors. This will have implications for industries that use autonomous robots at scale (eg, warehouses, traffic, and monitoring).

(2) The topic of environment optimization is classically under-studied. Our initial set of publications has elucidated the importance and value of considering the environment as a decision variable, and has brought this topic to the forefront. In particular, we have provided strong theoretical guarantees on agent performance (under optimized environments), which will encourage uptake of the idea in the community at large across application sectors.