Robot arm learns in simulation to aim a camera for flow measurements in a water tunnel

Researchers led by Chang Rajani have trained a reinforcement learning agent, described in an arXiv paper, to move a camera on a six-axis robot arm so it finds, frames and resolves the wake behind a hydrofoil in a water tunnel for particle image velocimetry, which measures flow from the motion of tracer particles. The agent sees the scene only through the velocity field and its uncertainty.

It was trained entirely in a photorealistic simulation of the water tunnel, with wakes from fluid dynamics simulations. In the real tunnel, without further training, it reliably located the wakes of randomly placed hydrofoils and resolved them at higher resolution and with lower error than a fixed camera, in about half the time of a full scan.