Merging Reaction to Cognition: A Hybrid Cognitive Strategy for Odor Source Localization in Natural Environments

ABSTRACT Chemical pollutants released into the environment are transported by turbulent flows, generating complex and intermittent plume structures that threaten ecosystems and human health. Rapid localization of emission sources is critical, and field robots equipped with chemical sensors provide a viable means to perform this task. Nevertheless, inferring source location from sensor readings remains difficult due to sparse detections and the absence of reliable concentration gradients. Existing approaches broadly fall into two paradigms. Bioinspired strategies rely on reactive behaviors triggered by chemical detections, such as surge‐casting, offering efficiency but requiring scenario‐specific tuning. Cognitive strategies integrate observations into a probabilistic belief over source location to guide exploration. While more robust, these methods suffer from excessive exploration and strong dependence on belief accuracy. The Fast‐Cognitive algorithm alleviated the computational burden of cognitive approaches but preserved their fundamental limitations. Previous Markov chain analysis revealed that source‐directed motions occur approximately twice as often following odor detections, indicating that reactive behaviors naturally emerge within cognitive frameworks. This work proposes a targeted extension of Fast‐Cognitive in which the continuous uncertainty‐based mixing of exploratory and exploitative motion is replaced by a detection‐triggered switch between them, prioritizing source proximity over information gain whenever chemical evidence is present. Both behaviors remain entirely parameterized by the belief, so the reactive layer requires no scenario‐specific tuning. The approach is validated through simulations under three turbulence conditions and field experiments with an autonomous surface vehicle in the Mondego River, Portugal. In simulation, the hybrid mechanism reduces traveled distance by 26%, rising to 39% in the most distorted scenario, where the success rate improves from 60.5% to 94.5%. In the field, the unmodified behavioral parameter set from simulation localized the source in six of seven trials with 3.2 m average error, providing preliminary evidence of transfer under the conditions tested.