Interfacial locomotion

Differentiable physics for wave-driven robots

Riding the interface: differentiable physics for wave-driven locomotion

A small robot sitting on water can propel itself forward without paddles, fins, or jets: by vibrating. The vibration radiates surface waves, and if those waves are directionally asymmetric, their momentum imbalance generates a net thrust. The SurferBot (Rhee et al., 2022) demonstrated this experimentally; our work builds a simulator where every design choice can be optimized directly.


At an air–water interface, surface tension, gravity waves, and added-mass effects govern the interface dynamics. Performance depends on coupled choices: body shape, mass distribution, motor location, drive frequency, waveform, and fluid properties, all of which are expensive to explore by experiment. We model the robot as a buoyant, possibly flexible body constrained to the interface and driven by a time-varying actuator, with the surrounding fluid described by an interface-resolving small-amplitude free-surface theory (Benham et al., 2024). The simulator is differentiable with respect to all design parameters $\theta$: state updates use linear and nonlinear solves $A(\theta)\,y=b(\theta)$ with custom reverse-mode rules, so $\nabla_\theta \mathcal{L}$ follows from two linear solves (forward and adjoint) per time step, keeping memory bounded and gradients stable across the full trajectory.

With these gradients, multi-start optimization explores hull geometries, actuator placements, and drive waveforms; Bayesian optimization handles global search under power-budget and manufacturability constraints.

References

2024

  1. Arxiv
    On wave-driven propulsion
    Graham P. Benham, Olivier Devauchelle,  and  Stuart J. Thomson
    Journal of Fluid Mechanics, 2024

2022

  1. SurferBot: a wave-propelled aquatic vibrobot
    Eugene Rhee, Robert Hunt, Stuart J Thomson, and 1 more author
    Bioinspiration & Biomimetics, Jul 2022