Inspire Robots and Universities Open-Source HandEdit With 200M Image Samples

Inspire Robots, together with Fudan University, Shanghai Jiao Tong University and other academic and research institutions, has officially open-sourced the HandEdit dataset and benchmark.

Focused on first-person human-to-dexterous-hand image editing, the release offers large-scale data resources and a standardized evaluation basis for dexterous manipulation learning.

Fine-grained tasks such as grasping, rotating, pressing and assembly place high demands on a dexterous hand's grip, perception and control precision. Yet data accumulation for robot manipulation still lags well behind the rapid advance of hardware.

Conventional collection relies on teleoperation, motion capture or repeated real-robot execution, which is costly and slow. Because dexterous hands differ in finger count, joint structure, size and range of motion, a single dataset rarely transfers across robot platforms.

Meanwhile, the internet already holds a vast supply of human-hand manipulation videos, but structural and degree-of-freedom differences between human and robotic hands make that information hard to convert into data a dexterous hand can learn from and execute.

Solving this data-adaptation problem is key to further progress in dexterous manipulation learning, and HandEdit explores exactly that question.

HandEdit studies first-person human-to-robot dexterous-hand image editing: keeping the original scene and action intent intact while replacing the human hand or arm with a specified robot configuration.

It is not a simple visual effect. The edit must fully remove the original hand while preserving object state, contact relations and background, and the generated robot must match the structure, materials and joint morphology defined by the target URDF — requiring scene understanding, motion preservation, configuration consistency and hand-object interaction to be handled together.

Inspire Robots' DFX and G2 series dexterous hands are included in this data conversion pipeline as key configurations.

In a coffee-machine manipulation scenario, the DFX series dexterous hand used HandEdit as a bridge to complete the transition from human demonstration to robot execution.

The case illustrates how the data translation pipeline operates on a real manipulation task.

The HandEdit generation framework preserves the original scene and action intent while automatically mapping human hand trajectories onto the DFX hand's native joint motion.

In tests such as blender grasping, DFX uses its own joint limits and parameters to output adapted rotation angles, contact positions and finger flexion in real time, keeping every frame within the target robot's physically reachable range.

HandEdit has so far built more than 200 million image editing samples, covering 26 URDF configurations — 13 standalone dexterous hands and 13 integrated arm-hand structures.

Spanning over 600 scenes, more than 1,100 objects and over 400 task categories, with a unified evaluation system, it provides a larger-scale data foundation for robot manipulation learning and a new technical path for turning human manipulation data into robot data.