Modeling one-shot interceptions in fruit-catching fish

This study explores the behavior of interception in fruit-catching fish, specifically focusing on the characid fish Brycon guatemalensis. By applying stochastic optimal control theory and reinforcement learning, researchers aim to understand the one-shot interception process, which is a universal pattern in animal behavior with potential industrial applications.

In the research, an agent acts to intercept a moving target, making decisions about wait time before launch and heading angle based on a noisy estimate of the target's initial state. In a simplified scenario where launch velocity and actuated control remain constant, the optimal wait time is influenced by noise in both the target measurement and control actuation, as well as the agent's energy conservation efforts.

The researchers derive an upper bound for the heading angle variance, which scales proportionally to the ratio of the non-vanishing optimal interception error to the initial planar separation. Additionally, they develop a deep reinforcement learning approach that enables the neural network representing the fish to learn the value function for interception, how to aim, and finally, the optimal wait time.

While this study is centered on the specific problem of fruit-catching in fish, the findings could also be applied to other scenarios with limited time for feedback or corrections, such as perching, landing, and docking in other contexts.