MARL-AUV for Spatiotemporal Mapping

ICRA-26 submission under review

This research focuses on developing collaborative multi-agent learning algorithms and adaptive control strategies for autonomous underwater vehicles to efficiently perform long-term spatiotemporal mapping in dynamic, resource-constrained settings. The work integrates data-driven models with energy-aware decision-making to improve mission endurance and mapping accuracy in complex real-world environments. We have one paper published on arXiv and currently under review at ICRA 2026, demonstrating how multi-agent reinforcement learning combined with Gaussian process regression enables coordinated AUV fleets to map river plumes over multiple days while balancing estimation accuracy and energy efficiency.

Current extensions include: (1) extending our work to 3D, (2) hardware deployment by interfacing our RL policy with low-level vehicle control, (3) replacing Gaussian process regression with neural operators for more scalable spatiotemporal modeling, and (4) theoretical work on multi-agent communication strategies using Jumanji and MAVA environments.

📽️ Watch 3 agents in action!