AI System Uses Belief Updating to Locate and Characterize Hidden Sources

Finding a hidden source in a complex, shifting environment using only a few measurements is a challenge that spans robotics, environmental monitoring, and search-and-rescue operations. A new research approach tackles this problem by teaching a mobile sensor to actively hunt down the source—and learn what it is—at the same time.

The Core Problem#

In inverse source localization, a sensor must infer where a source lies and what its properties are while moving through space and gathering noisy readings. The challenge intensifies when the sensor has incomplete information about the environment (partial observability), the field changes over time, and the number of measurements is limited. Traditional methods struggle because they don't adapt sensing strategy in real time based on what's been learned so far.

A Unified Framework#

The proposed system, called ATT-PFRL, weaves together three essential components:

• Better belief tracking: An attention-based particle filter improves how the system maintains and updates its internal estimate (posterior) of where the source likely is. This happens through careful resampling when confidence drops, targeted attention to key features, and a technique called Metropolis-Hastings rejuvenation that keeps the estimate from degrading.

• Smart stopping rules: Rather than collecting data blindly until a deadline, the system knows when it has enough information. Belief contraction—how tightly clustered the system's uncertainty estimates are—signals both when to stop searching and serves as a built-in reward signal, eliminating the need to tell the agent the true distance to the source.

• Reinforcement learning: The agent learns to choose sensor positions that shrink its uncertainty most efficiently, without explicit coaching about proximity to the target.

Testing and Results#

The team validated ATT-PFRL across seven different types of fields (diffusion, acoustic, electromagnetic, and others), including scenarios where the source moved unexpectedly and cases where the test environment differed from training. In all cases, the system outperformed planning-based and hybrid RL+Bayesian baselines on three key metrics: task completion rate, convergence speed, and localization accuracy. Separate experiments using fixed sensor paths further showed gains in the quality of the filtering step itself, isolating the inference improvement from the learning component.