Neural mechanisms for reward-modulated vector learning and navigation: from social insects to embodied agents

摘要: Despite their small size, insect brains are able to produce robust and efficient navigation in complex environments. Specifically social insects, such as ants bees, these navigational capabilities guided by orientation directing vectors generated a process called path integration. During this process, they integrate compass odometric cues estimate current location vector, home vector for guiding them back on straight path. They further acquire retrieve integration-based memories anchored globally the nest or visual landmarks. Although existing computational models reproduced similar behaviors, largely neglected evidence possible neural substrates underlying behavior. Therefore, we present here model of mechanisms modular closed-loop control - enabling embodied agents. The consists integration mechanism, reward-modulated global local learning, random search, action selection. mechanism integrates compute vectorial representation agent's activity patterns circular arrays. A learning rule enables acquisition associating food reward with state. motor output is computed based combination exploration. In simulation, show that enable homing localization, even presence external sensory noise. proposed rules lead goal-directed route formation performed under realistic conditions. This provides an explanation for, how view-based strategies Consequently, provide novel approach simulated agent linking behavioral observations substrates.