Tsinghua University researchers introduce benchmark dataset for dexterous robotic hands

A research team at Tsinghua University has developed HandsOn, a benchmark designed to evaluate dexterous robotic hands, 36kr reported. The dataset includes 66 specialized tasks alongside 18,000 high-quality human operation demonstrations. The evaluation suite aims to address a gap in robotic dexterity assessment by establishing a standardized, high-difficulty benchmark for bimanual manipulation. [1]

According to 36kr, earlier robotic manipulation benchmarks were primarily constructed around simple grippers rather than multi-fingered hands. The new dataset leverages first-person human manipulation data to capture how human hands grasp objects, utilize tools, and carry out delicate contact operations. [1]

Humanoid hands remain complex mechanical subsystems, as a biological human hand possesses approximately 27 degrees of freedom and thousands of tactile receptors. Standard robotic manipulation approaches historically relied on simple two-finger end-effectors that often lacked pressure sensors for calibrating force, making fine contact tasks difficult to benchmark. [3]