Learning based approaches to practical challenges in multi-agent active search

Interactive decision making is essential for the functioning of autonomous agents in both software and embodied applications. Typically, agents interact in a multi-agent environment with the goal of fulfilling individual or shared objectives. In this thesis, we study the multi-agent adaptive decision making problem in the framework of Multi-Agent Active Search (MAAS) with a focus on applications like search and rescue, wildlife patrolling or environment monitoring with multi-robot teams.

MAAS involves a team of robots (agents) deciding when and where to gather information about their surroundings, conditioned on their past observations, in order to estimate the presence and position of different objects of interest (OOIs) or targets. Agents communicate with each other asynchronously, without relying on a central controller to coordinate the agents’ interactions. Realistically, inter-agent communications may be unreliable, and robots in the wild have to deal with noisy observations and stochastic environment dynamics. Our setup, described in Chapter 1, formalizes MAAS with practical models of real-world sensing, noise, and communication constraints for aerial and ground robots.

Part I of this thesis studies the benefits of non-myopic lookahead decision making in MAAS with Thompson sampling and Monte Carlo Tree Search. Additionally, we consider a multi-objective pareto-optimization setup for cost-aware active search, highlighting the challenges due to partial observability, decentralized multi-agent decision making, and computational complexity of combinatorial state and action spaces. In Part II, we focus on the practical challenges due to observation noise and dynamic targets in multi-agent active search and tracking. Our proposed algorithms using Bayesian filtering in these settings em- pirically demonstrate the importance of uncertainty modeling for inference and decision making. Part III shifts focus to generative models for decision making, particularly denoising diffusion sampling for lookahead MAAS with observation noise. Finally, we discuss the applicability and limitations of these methods in the context of multi-agent decision making in robotics and other applications with similar real world constraints.