A self-evolving scientific agent discovers an interpretable fluid controller

Researchers report a self-evolving scientific-agent workflow that generates, tests, diagnoses, and revises control code inside a physics simulation. The demonstration uses a nonlinear fluid-structure problem: steering an underactuated two-joint robotic fish using only joint angular accelerations. Starting from a biased propulsive policy, the agent developed a readable controller combining traveling-wave propulsion, target guidance, yaw-rate feedback, tail curvature, and adaptive cadence. The final policy reportedly generalized to unseen static targets and curved pursuit trajectories without retraining or target-specific branches. This is a simulation preprint, not proof that autonomous agents can reliably discover controllers for real machines. Its significance is methodological: it combines code-generating agents with auditable physical reasoning rather than treating the resulting controller as an opaque neural policy.

Key details: arXiv:2606.08405, June 7, 2026, Self-evolving scientific agent, Fluid-structure control simulation, Interpretable generated controller, Generalized without retraining.

Continue swiping for more AI Brief stories.