HydroShear gives robots a sense of touch with 93% success rate

HydroShear, developed by researchers Mani Nambi and Nima Fazeli from Amazon and the University of Michigan, is a physics-based simulator that bridges the 'tactile reality gap' for robotic manipulation. It introduces path-dependent force tracking within hydroelastic contact models, accurately simulating how shear forces accumulate over a soft sensor membrane during contact. Unlike prior methods that oversimplify force dynamics, HydroShear remembers each contact point's motion history—computing realistic friction, slipping, and elastomer deformations as objects slide, tilt, or roll against the sensor. This allows robots to learn contact-rich tasks purely in simulation, with policies that transfer seamlessly to the real world without modification.

In real-world tests using a Franka robot arm equipped with GelSight Mini tactile sensors, HydroShear achieved a 93% average success rate across four challenging manipulation tasks—grasping, tool use, and object rotation—compared to only 34% for TacSL and 58-61% for FOTS. The simulator is GPU-parallelizable, enabling large-scale reinforcement learning training that was previously impossible due to the slowness of finite-element methods. This breakthrough significantly reduces the need for expensive real-world data collection and trial-and-error learning, opening the door for robots to master delicate tasks in warehouse automation, surgical assistance, and more.