Autonomy at the Edge of Capabilities: Uncertain Environments, Limited Resources, Difficult Missions

Overview

Fast, successful, and efficient planning is a core challenge of high-level autonomy in complex environments. The obstacles are seemingly insurmountable. Individual agents often face constraints in terms of resource and compute availability​, limited sensing and communication capabilities, and lack of a priori knowledge about the operating environment. Planning for large teams is burdened by either curse of dimensionality or complex organizational patterns of decentralization. As a result, standard machine learning methods are largely infeasible — for instance, due to lack of training data or the size of the state space — while human-driven solutions or simple heuristics often produce vastly suboptimal plans. The purpose of this talk is to propose a middle road. We will consider three broad problems in planning: budget-constrained teams, task-aware data collection, and time-optimal multi-target planning. Across two domains — infrastructure maintenance and maritime autonomy — we show that understanding the structure of agent interactions and the interplay between environment and mission progress is key in developing meaningful, computationally tractable policies. Consequently, our strategies combine machine learning reasoning with high-level structure-driven abstraction and mission decomposition. Early empirical work demonstrates that such an approach greatly outperforms existing benchmarks while retaining the capability to operate at impressively large scales.