Reproduce “Dense Force Estimation with an Event-based Optical Tactile Sensor” on OpenResearch

About this paper

Agis Politis, René Zurbrügg, Valentina Cavinato

Researchers developed a framework for dense 3D force field reconstruction using an event-based optical tactile sensor, which leverages a neural network to infer normal deformation from sparse event data. The system operates at 100 Hz, achieving mean absolute errors of 0.14 N (shear X), 0.10 N (shear Y), and 0.93 N (normal Z) for force estimation.

Key insights

• Sparse event streams, though not directly encoding depth, contain implicit global marker pattern information that can be learned by a neural network to infer dense normal deformations.
• A contact-aware loss function effectively improves the accuracy of normal deformation estimation, particularly across diverse contact geometries and varying indentation depths.
• Combining high-temporal resolution event-based tracking with learned depth estimation and a physically-grounded iFEM model enables dense 3D force feedback at rates significantly higher than traditional vision-based methods.

Summary by alphaXiv.