RSS 2026 HydroShear

HydroShear: Hydroelastic Shear Simulation for Tactile Sim-to-Real Reinforcement Learning

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: RL · paper #155 Authors: An Thanh Dang, Jayjun Lee, Mustafa Mukadam, X. Alice Wu, Bernadette Bucher, Manikantan Nambi, Nima Fazeli (University of Michigan · Amazon Industrial Robotics) arXiv: 2603.00446 · program page

Summary compiled from the arXiv paper (v1); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Figure 1: The four real-world manipulation tasks used to validate HydroShear, shown in simulation (top row) and on the real robot (bottom row): (a) peg insertion under in-hand pose uncertainty, (b) bin packing requiring multi-object contact, (c) book shelving with gravity orthogonal to the insertion axis, and (d) drawer pulling that tests precise, minimal gripper-force modulation under slip. Each task highlights a different tactile-shear challenge.

Problem

Tactile sim-to-real transfer for contact-rich tasks is hard because existing tactile simulators emphasize image-rendering quality while modeling force and shear too simplistically, producing a large sim-to-real gap for dexterous manipulation that depends on extrinsic contact and slippage.

Method

HydroShear is a non-holonomic hydroelastic tactile simulator that models (a) stick–slip transitions, (b) path-dependent force and shear build-up, and (c) full SE(3) object–sensor interactions. It extends hydroelastic contact models using Signed Distance Functions (SDFs) to track displacements of an indenter's on-surface points during contact with the sensor membrane, generating physics-based, computationally efficient force fields from arbitrary watertight geometries while remaining agnostic to the underlying physics engine. Policies are trained with PPO in simulation on GelSight Mini sensors and deployed zero-shot on the real robot, with a digital-twin calibration step against real marker displacements.

Results

With GelSight Minis, HydroShear reproduces real tactile shear more faithfully than prior methods, enabling zero-shot sim-to-real RL transfer across the four tasks (peg insertion, bin packing, book shelving, drawer pulling). It achieves a 93% average success rate, versus 34% for policies trained on tactile images and 58%–61% for alternative shear-simulation methods.

Significance

Advances high-fidelity tactile simulation as a route to zero-shot sim-to-real RL for fine, force-sensitive manipulation, relevant to the RL and Review-Dexterous-Manipulation threads.

← Back to RSS 2026 survey · RSS-2026-Papers · Home