Survey maps learning methods for dexterous robotic hand control
Learning Dexterous Manipulation: Hands, Simulators, Training, and Evaluation
Robotics
Summary
Changing the position of objects with robotic hands without putting them down is very hard. The authors review 113 methods that use learning to solve this, examining what robot hands, simulators, and training approaches they use. Most methods rely on a popular way: training with detailed internal data in a fast computer simulation and then teaching a vision-only model to work on real robots. Another group learns from human demonstrations using vision and language. They found growth in studies, mostly on grasping and using two hands, but many robot hands and details remain untested or unreported.
What this means in practice
- •For robotics engineers: Choose and develop learning methods with awareness of current dominant simulators, hands, and evaluation gaps for dexterous robotic manipulation tasks.
- •For robotics hardware designers: Identify underexplored robotic hand designs to improve hardware-software integration for dexterous object manipulation.
A survey. It maps existing work.
Authors
Luai Abuelsamen
Abstract
A dexterous hand can change an object's pose without setting it down. This survey covers 113 learning-based methods that attempt it, the hands they run on, the simulators they train in, and how they report results. The field has largely converged on one recipe: reinforcement learning on privileged state in a GPU simulator, distilled into a vision-only student for transfer to hardware. Almost half the corpus trains that way, nearly all of it in one simulator, whose own successors are barely used. A second and faster-growing branch learns from teleoperated demonstration instead, through vision-language-action models, diffusion policies, and systems whose contribution is the collection rig itself. Published work has tripled since 2022. Grasping remains the largest task family and bimanual coordination the fastest-rising, and almost every experiment runs on one of four hands. What the survey adds comes from reading released code and vendor specifications beside the papers. Eight hands that can be bought or built from published designs appear in no method paper here. Where the depth a hand sinks into the object it holds is reported at all, it is not measured in the physics that produced the motion. Section VII proposes a protocol for the unreported quantities; the body gives the counts behind every claim above.