Study on the deployment and design of tactile sensors in dexterous manipulation

Tactile sensing is essential for enhancing humanoid perception and enabling learning-based dexterous manipulation by providing real-time force, slip, and grip feedback for human-like interaction. Yet large-scale adoption remains constrained by dual challenges in hardware and deployment strategy: commercial sensors emphasize single-modal resolution with expensive components and complex fabrication, while even low-cost options face prohibitive wiring, integration, and computational overhead for full-hand arrays in DRL settings. We address these issues by (1) introducing Bio-Skin, a cost-efficient multi-modal tactile sensor, and (2) systematically identifying sensor configurations that balance DRL performance with hardware efficiency. Bio-Skin employs single-axis Hall sensors for normal force, bar-shaped piezoresistors for 2D shear, and a thermistor-heater pair embedded in silicone for temperature sensing and thermostatic regulation, with a cross-reference scheme improving signal fidelity in complex electromagnetic environments. Its multilayer design enables rapid sequential fabrication, and after calibration it matches commercial signal-to-range ratio, sampling rate, and range at one-tenth the cost. We validate Bio-Skin on an Allegro hand for grasping and material recognition. For deployment, we study how sensor count and placement influence DRL performance: beginning with 92 sensors, we show that 21 retain over 93% performance. An active-learning pipeline using Gaussian Process Regression identifies high-uncertainty layouts for policy refinement, while Lasso regression quantifies each sensor’s importance, revealing both essential and low/negative-value units. Across five tasks (Block, Egg, Pen, Pentagonal Prism, Capsule), a reduced high-importance subset preserves over 90% performance. This work therefore provides both scalable tactile hardware and principled deployment guidelines, enabling cost-effective, large-scale dexterous hands under diverse sensor budgets.