Chinese robotics startup AGILINK has demonstrated its robot hand performing basic embroidery techniques using silk thread. The demonstration involved separating a silk thread into 16 finer strands, threading a needle, stretching fabric across a hoop, and making initial stitches.
The firm’s OmniHand 3 Ultra robot hand used coordinated finger movements to maintain tension while separating and handling the delicate strands.
The demonstration highlights the challenges of using robotic hands for precisely manipulate thin, flexible materials.
Master silk embroidery
AGILINK, a spin-off of Chinese robotics company AgiBot, has demonstrated its OmniHand 3 Ultra dexterous robot hand performing techniques used in traditional Suzhou embroidery. The demonstration involved training the robotic hand to split silk thread, thread a needle, stretch fabric across an embroidery frame, twist thread, and make stitches.
Suzhou embroidery requires precise handling of silk, which can bend, slip, and change shape under small changes in applied force. In the demonstration, embroidery master Fu Xianghong taught the robot how to manipulate the material—one of the key steps involved separating a main silk thread into 16 finer strands. The robot used one finger to separate the strands while its other fingers maintained tension.
The hand also had to keep its fingertips in contact with the silk while allowing the thread to rotate during twisting. Threading the needle and stitching required additional control over position, force, and finger coordination.
“Silk has no rigid shape. The force of the fingers decides it. That makes embroidery a test of hardware and control together: a fully direct-drive hand with 21 active degrees of freedom, and a visual-tactile sensor in every fingertip that reads contact from surface deformation,” reads the video description posted by the firm.
AGILINK said the training process combined teleoperation data with reinforcement learning. In the final task, the robot performed a continuous motion that involved twisting and then picking up the thread. Because the thread landed in a different position each time, the task required the robot to adapt its movements rather than follow a fixed sequence.
Hand gains touch
The OmniHand 3 Ultra is about the size of a human hand and weighs 1.9 pounds (630 grams). It uses a direct-drive architecture, with motors connected directly to the joints rather than using tendons or intermediate transmission mechanisms.
The hand can reportedly hold 6.6 pounds (3 kilograms) steadily and lift to 17.6 pounds (8 kilograms). It can open and close in about 0.3 seconds and has a positional repeatability of about 0.008 inches (0.2 millimeters).
Touch sensing is another part of the system. Each fingertip uses a vision-based sensor that detects changes in the deformation of its contact surface. According to AGILINK, the sensors can measure force in three dimensions and detect deformations as small as 0.08 millimeters. The palm also incorporates a 300-point contact sensor.
The combination of mechanical design, tactile sensing, and software control allows the hand to respond to materials that are difficult to manipulate using rigid robotic grippers. Silk presents a particularly demanding case because its shape and position can change during contact.
AGILINK used the embroidery demonstration to show how its dexterous hand can perform tasks requiring continuous adjustments to force and finger position. The approach differs from robotic systems designed primarily for repetitive industrial operations, where objects and movements can be more predictable.
Other companies are also developing dexterous robotic hands with different design priorities. LinkerBot’s Linker Hand L30 Pro uses five tendon-driven fingers and is designed for speed and repeatable movements. Boston Dynamics has taken a different approach with its latest Atlas hand, which uses four digits and is aimed at industrial applications such as handling drills and welding tools.