HapticNet: Building and Evaluating a Haptic Knowledge Base with an Agentic AI Workflow

Touch remains one of the least digitized human senses. While vision, language, and audio have been transformed by internet-scale data, haptics remains constrained by small-scale, lab-collected signals rather than large, structured repositories of material knowledge.

We present HapticNet, a structured, evidence-grounded dataset of 1,230 materials—the largest of its kind—organized hierarchically and annotated with seven quantitative haptic properties spanning compliance, roughness, friction, and thermal response. Each entry is represented as a normalized, provenance-aware record, capturing values, units, and measurement conditions.

To support rigorous evaluation, we introduce HapticNetEval, a human-verified benchmark comprising 1K queries and 5K source-grounded value–condition records. The benchmark provides deterministic metrics for correctness and grounding, complemented by calibrated judge-based evaluation.

Finally, we demonstrate Haptic Inference, a downstream pipeline that maps visual inputs to material identities and retrieves corresponding haptic properties, enabling reusable haptic assets for simulation, rendering, and remote touch applications.

M.Sc. student under the supervision of Prof. Lihi Zelnik – Manor.