Distill-Belief: Closed-Loop Inverse Source Localiza... | AI Research

Distill-Belief: Closed-Loop Inverse Source Localization and Characterization in Physical Fields
This research addresses the challenge of autonomous robots or drones tasked with locating sources—such as gas leaks, pollutants, or radiation—in complex physical environments. These agents must make real-time decisions about where to take measurements under strict time and energy constraints. The core difficulty is that the agent needs to maintain a "belief" about the source's location and characteristics while simultaneously deciding where to move. Traditional methods often struggle because they either rely on computationally expensive Bayesian inference that is too slow for real-time use or use simplified models that can lead to "reward hacking," where the agent exploits errors in its own logic to claim success without actually finding the source.

The Teacher-Student Framework

To solve this, the authors introduce a teacher-student architecture that separates scientific accuracy from operational efficiency. The "teacher" is a Bayes-correct particle filter that maintains a high-fidelity, statistically accurate map of the environment's parameters. This teacher is used only during the training phase to provide a reliable signal for the agent to learn from. The "student" is a compact, efficient model that learns to mimic the teacher’s belief. By distilling the teacher’s complex posterior into simple, fast-to-compute statistics, the student enables the agent to make decisions in constant time during deployment.

Preventing Reward Hacking

A major contribution of this work is how it handles the learning process. In many systems, if an agent uses its own internal belief model to define its "reward" for success, it may learn to manipulate that model to report high confidence even when it is wrong. Distill-Belief prevents this by ensuring that the intrinsic reward—the signal that tells the agent it is making progress—is calculated exclusively by the teacher. Because the teacher is not influenced by the agent's policy, the agent cannot "hack" the reward; it must genuinely reduce uncertainty about the source to receive a positive signal.

Deployment and Performance

At the time of deployment, the teacher is discarded entirely. The agent relies solely on the student’s distilled belief statistics to navigate and decide when to stop sensing. This allows the system to operate with a fixed, low computational cost regardless of the complexity of the environment. Experiments across seven different field types and two stress tests demonstrate that this approach consistently outperforms existing methods. It reduces the cost of sensing, improves the accuracy of source localization, and ensures that the agent stops only when it has achieved a verified level of certainty, effectively balancing the trade-off between mission time and estimation quality.