Stein Variational Ergodic Search

Overview

In continuous spaces there are infinitely many ways to cover an area, and committing to one can leave a robot stuck when conditions change. This work frames ergodic search as probabilistic inference and uses Stein variational methods to approximate a posterior distribution over ergodic trajectories in parallel.

How it works

• Ergodic search is reformulated as a probabilistic inference problem over coverage trajectories.
• Stein variational gradient descent optimizes many trajectories in parallel to approximate the posterior, giving a diverse set of feasible coverage strategies.
• Embedded in model-predictive control, the robot can adapt its exploration strategy online.

Experiments

Simulated and physical experiments show adaptability and diversity in exploration strategies online.

My role

Co-author. [Your contribution: e.g. physical experiments, drone platform.]