SWiFT framework lets robotic fish hold station in turbulent flow without sensors

A team led by Xiaozhu Lin (with co-authors from Peking University and others) introduced SWiFT (Swimming With Flow Toolbox), a framework that enables a body and/or caudal fin (BCF) robotic fish to approach and hold position in unknown turbulent background flows. The core innovation is training a station-holding policy via reinforcement learning (RL) using only egocentric feedback—no explicit flow sensing. The framework combines a free-swimming flow-tank experimental setup, a highly efficient CFD-based simulator that ensures physical consistency, and a systematic sim-to-real transfer pipeline. Benchmarking against state-of-the-art methods, SWiFT achieved substantial improvements across all metrics, most notably root-mean-square error (RMSE) of distance.

The success of egocentric station-holding in unknown turbulence mirrors the biological phenomenon of rheotaxis, where fish maintain position using only proprioceptive sensing. This not only reduces sensor cost and complexity but also opens the door to real-world deployment in autonomous underwater vehicles (AUVs). The SWiFT framework is positioned as a foundation for tackling more complex swimming tasks, such as schooling, obstacle avoidance, or searching in chaotic currents. The paper includes 20 pages and 18 figures, underscoring the depth of experimental validation. Published on arXiv in July 2026, this work represents a significant leap for bio-inspired underwater robotics.