Decoupling what, how, and when for observing decision-making context in autonomous robots

This paper presents an approach that decouples what to observe, how to observe it, and when observations are required for decision-making in autonomous robots. Situation awareness is essential for efficient and reliable autonomous robot operation, but despite advances toward parallelizing perception and action, key challenges remain in making perception aware of the current context and ensuring observability during action execution. To address this, we explicitly model the decision-making context and identify it with dedicated observers running in parallel to the task execution. Observer behaviors are implemented as behavior trees and coordinated by a centralized context manager. We validate the approach on a mobile manipulator that performs autonomous machine-tending tasks in a real medical laboratory, as well as during a public trade fair. Experimental results show that context-driven observers can robustly identify decision-making context without interfering with ongoing actions. Furthermore, parallelized perception leads to substantial runtime improvements, achieving overall task time savings of up to $24\%$ . These findings demonstrate that explicit context modeling and observer-based perception parallelization enhance the stability and efficiency of existing robotic execution architectures.