Multi-Agent Coordination
Scalable Decentralized Coordination for Complex Multi-Robot Systems
This project introduces MARVEL (Multi-Agent Reinforcement Learning for Constrained Field-of-View Multi-Robot Exploration), a framework for high-performance, decentralized coordination in large-scale environments. By leveraging Graph Attention mechanisms, MARVEL enables robot teams to reason about teammate intent and spatial dependencies under restricted sensing constraints. Our approach focuses on information-theoretic action pruning to optimize coverage and mission efficiency, facilitating complex collaborative maneuvers in completely unknown topographies without a central controller. Its follow-up, Dec-MARVEL, drops inter-robot communication altogether, coordinating on opportunistically observed teammate trajectories while holding every robot to a travel budget that guarantees a return to base.
MARLGraph Neural NetworksInformation-Theoretic PlanningMulti-Robot Coordination
Contributions
What this work contributed, drawn from the papers below.
-
01
Exploration under a constrained field of view
Most exploration planners assume omnidirectional sensors such as LiDAR, which payload-limited drones cannot carry. MARVEL targets robots whose only option is a lightweight directional camera, so the planner must decide both where a robot moves and where its sensor points.
MARVEL, ICRA 2025 -
02
Frontier and orientation fusion over graph attention
A neural framework fuses frontier features with sensor-orientation features inside graph attention networks, producing a collaborative decentralized policy trained with multi-agent reinforcement learning.
MARVEL, ICRA 2025 -
03
Information-driven action pruning
Viewpoint planning opens up a very large action space. An information-driven pruning strategy narrows the candidate viewpoints so the learned policy stays tractable.
MARVEL, ICRA 2025 -
04
Transfer across team sizes and sensor configurations
Learned policies adapt to different team sizes, fields of view, and sensor ranges without any additional training.
MARVEL, ICRA 2025 -
05
Validated at scale and on real drones
MARVEL outperforms state-of-the-art exploration planners across multiple metrics, generalises to environments up to 90 m by 90 m, and was deployed successfully on a team of real drone hardware.
MARVEL, ICRA 2025 -
06
Coordination without a radio link
Wireless links degrade in exactly the environments worth exploring. Dec-MARVEL removes inter-robot communication entirely and coordinates on opportunistic sightings instead: a teammate trajectory that happens to fall inside a robot's field of view tells it which regions have just been covered.
Dec-MARVEL, arXiv 2026 -
07
Return-feasible planning under a travel budget
Every candidate move is checked against the shortest path home plus a safety margin, so a robot can never explore itself out of range of its base. Training pairs a task-oriented privileged critic and phase-conditioned value heads with a budget curriculum that tightens as competence grows, cutting sensing overlap by 26.4% and travel to full coverage by 17.7% against the strongest communication-free baseline.
Dec-MARVEL, arXiv 2026 -
08
Decentralized scheduling for heterogeneous coalitions
Tasks that can only begin once every required agent has arrived force robots to form and disband teams by capability, which creates idle waiting at task locations. A fully reinforcement-learned decentralized policy generalises across heterogeneous agents, where mixed-integer programming would need centralized scheduling and long optimization times.
Task Allocation, IEEE RA-L 2025 -
09
Constrained flashforward mechanism
To make complex cooperation learnable, a constrained flashforward mechanism guides and limits how agents explore, helping them predict the consequences of joining a coalition.
Task Allocation, IEEE RA-L 2025 -
10
Attention over cooperation and scheduling horizons
An attention mechanism reasons about both short-term cooperation and long-term scheduling dependencies, so agents reactively pick their next task and coalition to avoid wasted capabilities. The result closely matches or beats heuristic and mixed-integer baselines while running at least two orders of magnitude faster.
Task Allocation, IEEE RA-L 2025