Olfactory pursuit: catching a moving odor source in complex flows
A hybrid algorithm combines Infotaxis with predictive control to catch moving targets using intermittent scent signals.
A team of researchers including Maurizio Carbone has published a new paper on arXiv, 'Olfactory pursuit: catching a moving odor source in complex flows,' tackling a fundamental robotics challenge. The problem involves an autonomous agent locating and intercepting a moving target using only intermittent, delayed, and turbulent-mixed odor signals—a scenario akin to an animal tracking prey. The researchers framed this as a Partially Observable Markov Decision Process (POMDP), where the agent maintains a belief over the target's position and velocity. They computed quasi-optimal policies by solving the Bellman equation and benchmarked them against established strategies like Infotaxis.
They discovered that purely exploratory policies like Infotaxis work well when a target changes direction frequently but fail dramatically when it moves persistently. To solve this, the team introduced a novel, computationally efficient hybrid policy. This heuristic combines the information-gathering strength of Infotaxis with a 'greedy' predictive value function derived from a related, fully observable control problem. The result is a robust strategy that achieves near-optimal performance across all target persistence times and substantially outperforms older methods.
The new policy demonstrated strong robustness in complex simulations, including scenarios with continuous target motion, model mismatch, and more accurate plume dynamics. The research identifies predictive inference of target motion as the critical component for effective search in information-poor, dynamic environments. This work provides a general framework that could significantly enhance the capabilities of search-and-rescue robots, environmental monitoring drones, or any autonomous system operating based on sparse, noisy sensory data.
- Formulates odor tracking as a POMDP and solves for quasi-optimal policies via the Bellman equation, benchmarking against Infotaxis.
- Introduces a hybrid policy combining Infotaxis with a predictive 'greedy' controller, achieving near-optimal performance and beating purely exploratory methods.
- Demonstrates robustness in complex scenarios including continuous target motion and model mismatch, identifying predictive inference as key.
Why It Matters
Provides a robust framework for autonomous robots to perform critical tasks like search-and-rescue or pollution tracking using unreliable sensory data.