Direct-drive robotic hand · serves as: Grasping, Manipulation, Tactile sensor. · compatible with: Humanoid, Manipulator.
Which group OmniHand 3 Ultra-M belongs to and how it is built
A subcategory encompassing robotic hands with multiple degrees of freedom, designed to perform complex grasping, regrasping, manipulation, and object interaction tasks. Unlike simple grippers, dexterous hands enable more advanced finger coordination and a broader variety of grasp types.
Robotic hand whose joints are driven directly by motors (without tendons), enabling backdrivable operation.
Five-fingered robotic hands designed for a wide range of grasps and more anthropomorphic manipulation, commonly used in humanoid robots and manipulation research.
Basic physical properties of OmniHand 3 Ultra-M — dimensions, weight and materials
AGILINK provides a Python and C++ SDK, native simulation assets for NVIDIA Isaac and MuJoCo, and ROS 2 reference nodes.
The OmniHand 3 Ultra-M is the flagship anthropomorphic five-finger dexterous hand from AGILINK (an AGIBOT spin-off), built around direct drive at every joint. It provides 20 active degrees of freedom in a 630 g, human-scale body measuring 189.5 × 88.5 × 40 mm.
Direct drive delivers ±0.2 mm positioning repeatability across the full workspace and joint state read directly from each actuator encoder. Typical open/close time is 0.3 s. The hand stably grasps up to 3 kg and lifts up to 8 kg.
Tactile perception combines a vision-based 3D+6D fingertip sensor (0–30 N range, 0.005 N force resolution, 0.04 mm spatial resolution, 0.08 mm deformation accuracy) with a 3D palm tactile sensor featuring 300 taxels at 0.1 N resolution. Communication is over CAN FD, RS-485 and EtherCAT.
The OmniHand 3 Ultra-M targets continuous-duty operation (40–45 °C surface temperature under load with no throttling, 500,000 validated cycles), embodied-AI training, teleoperation, demonstration capture and bimanual contact-rich manipulation. AGILINK provides a Python and C++ SDK, native assets for NVIDIA Isaac and MuJoCo, and ROS 2 reference nodes.
