four problems at the water line.

Most legged robots ship one controller per regime: dry, mud, water. Switching between them is brittle — it depends on a clean classification of the terrain, which is exactly what falls apart in the intertidal zone.

We're treating substrate as a continuous variable. The same policy outputs a foot-placement strategy whether the ground is dry sand at low tide, partially saturated mud as the tide turns, or 15 cm of standing seawater on top of compacted clay.

Open questions: how to encode substrate state from on-board sensing alone, how to design a foot whose contact dynamics degrade gracefully across regimes, how to recover from an unexpected loss of traction without surface-fall.

Standard SLAM stacks assume a stable horizon, opaque ground, and a world that doesn't move on its own. None of that holds at the water line.

We're investigating perception pipelines that treat the water surface as a first-class feature — segmentation, reflection-aware feature matching, and learned priors on what a wave should look like versus a static object.

Open questions: how to fuse visual-inertial estimates when an imu is being shaken by wave impact, how to maintain a usable map across tidal cycles, how to do object permanence on a surface that genuinely is not permanent.

The boring half of the problem and the half that decides whether anything ships. Salt water is harder on machines than fresh, and the cycling between wet and dry is harder than continuous submersion.

We care about: sealed actuators rated for repeated splash and brief submersion, connectors and bearings that don't fail at the wet/dry seam, foot pads that survive abrasive mud and barnacle-grade rock.

Open questions: what duty cycle of wet/dry exposure does current commercial hardware actually survive in the field, where are the cheap reliability wins, and which platforms are worth designing around the failure modes instead of fighting them.

We think the wet boundary is too non-stationary for clean sim-to-real. Tidal phase, weather, season, sediment load, biofouling — the gap between any simulator and a real estuary is large and structural.

Our default is to train on site and evaluate on site. Simulation is for sanity checks and ablations, not for the primary loop.

Open questions: how to safely collect training data when a failure mode might mean a robot stuck in tide, how to design a fleet of cheap field probes that gather data faster than a single expensive platform can be tuned, and where the sim-to-real assumption actually does hold.