Tactile sim-to-real transfer often prioritizes realistic images while oversimplifying force and shear, leaving a large gap for dexterous manipulation. HydroShear, a non-holonomic hydroelastic tactile simulator, instead models stick-slip transitions, path-dependent force and shear buildup, and full SE(3) object-sensor interactions. The method extends hydroelastic contact models with signed distance functions to track displacements of on-surface points of an indenter as it interacts with the sensor membrane.
In experiments with GelSight Minis, the authors report that HydroShear reproduces real tactile shear more faithfully than existing methods. Reinforcement learning policies trained in simulation transferred zero-shot to four tasks—peg insertion, bin packing, book shelving, and drawer pulling—with a 93% average success rate. This outperformed policies trained on tactile images, which reached 34%, and policies trained with alternative shear simulation methods, which reached 58%–61%.
The paper, accepted to RSS 2026, argues that existing tactile simulators emphasize rendering quality at the expense of physical fidelity. By generating physics-based force fields from arbitrary watertight geometries, HydroShear offers a computationally efficient path toward more reliable tactile policy learning across physics engines.