Autonomous optical velocimetry via reinforcement learning: a proof-of-concept agentic autofocus
Year: 2026
Type: conference-paper
Abstract: Optical velocimetry is the state-of-the-art experimental technique in fluid dynamics offering a non-intrusive method to obtain high-resolution velocity fields in wind and water tunnel experiments. Despite exceptional theoretical, experimental, and hardware advances made over the last two decades by the particle image velocimetry (PIV) community, experimental setups continue to require time-consuming supervision from the expert human operator. Notably, navigating complex geometries with difficult optical access, avoiding light reflections from solid surfaces, losing focus, and escaping occlusion of important flow features require strenuous human input. Here, we propose a reinforcement learning (RL) framework forautonomous optical velocimetry experimentation. An RL agent is pre-trained in a virtual PIV environment and deployed on a robotic arm with a mounted PIV camera. We demonstrate the first proof-of-concept task where the learned control policy adjusts distance from the experimental setup for autofocus.
Institution: Empa
Language: English
Subfield: Computational Mechanics
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7205050879