Sensing flow gradients is necessary for learning autonomous underwater navigation
Published in Nature Communications, 2025
- Egocentric underwater navigation (using body-frame flow and target measurements) performs as well as geocentric sensing (using inertial-frame measurements) when supplemented with local flow-gradient information, either transverse or tangential flow gradients.
- Egocentric policies inherently obey rotational symmetry, maintaining high performance under arbitrary rotations of the flow field, whereas geocentric policies rapidly degrade when the flow alignment shifts.
- When generalizing to different flows (including from reduced order modle to actual flow), both egocentric and egocentric policies are robust in naviagting through vortical wakes.
- Analysis of the learned policies as vector fields reveals “preferred directions” (stable equilibria) that explain differing success rates upstream vs. downstream and the role of sensory ambiguities in decision-making.
Recommended citation: Jiao, Y., Hang, H., Merel, J. and Kanso, E. (2025). Sensing flow gradients is necessary for learning autonomous underwater navigation. https://doi.org/10.1038/s41467-025-58125-6