Magellan : a robust executive enabling long horizon multi-agent campaigns

Abstract

Autonomous mobile systems are being tasked to perform increasingly complex missions. These campaigns frequently involve the coordination of agents with continuous dynamics to achieve multiple goals over long horizons and often occur in hazardous environments that can change unpredictably. Campaigns are specified by hybrid activity plans, featuring both finite domain and real-valued variables. This thesis introduces a robust, centralized executive, Magellan, to facilitate the execution of long horizon hybrid campaigns. Previous hybrid execution approaches address the challenges of generating dynamically feasible trajectories that satisfy the goals of a higher level plan. However, these executives do not scale well to campaigns with long horizons or multiple agents. By leveraging the insights of previous work, Magellan robustly executes campaigns specified by hybrid activity plans, while monitoring, and adapting to, disturbances on-the-fly.

Our approach to robust hybrid execution hinges on three key innovations. First, we recognize that an executive must generate dynamically accurate trajectories to control continuous agents in real-time while also ensuring that all activities in the campaign plan can be achieved. Magellan address these competing needs using a receding horizon control strategy, only generating trajectories with sufficiently accurate dynamics, modeled in discrete time, over a limited horizon. Magellan avoids being myopic by reasoning over the full campaign plan, using a continuous time formulation with simplified dynamics, to guide the limited horizon trajectory. Magellan achieves a factor of 2 improvement in solution quality compared to the state-of-the-art. Second, hybrid activity plans are often full or partially grounded, however, grounded plans are brittle in the face of unforeseen disturbances during execution.

Magellan provides an algorithm for lifting a grounded hybrid activity plan to a flexibly executable plan that entails the same goals but is minimally constrained, allowing Magellan to adapt on-the-fly. Third, agents can deviate from plans during execution due to environmental uncertainty or actuator noise. Magellan monitors the state of the agents during execution and assesses whether constraints in the hybrid campaign plan have been violated. We present a procedure for monitoring both the finite domain and real-valued variables in a hybrid campaign plan.