Robotic hands are becoming increasingly capable of performing complex movements, but applying those capabilities reliably in industrial settings remains a challenge. At Agibot’s Innovation Day event on October 29, Agilink demonstrated how it is approaching the problem from two directions: testing its dexterous robotic hands in a factory environment and collecting the physical interaction data needed to train them.
According to Agilink, a proof-of-concept station equipped with its OmniHand was deployed on Wolong Electric’s production line in Shaoxing, where it conducted final performance tests on fully assembled robotic hands from the same series. The company said the station completed 72 consecutive operations, including loading materials, fastening screws, conducting self-checks, unloading, and returning to its starting position, without real-time human intervention.
A complete cycle took 8.5 minutes, compared with eight minutes for a skilled worker. Agilink reported a 98% task success rate and eight hours of continuous operation.
Fastening screws proved the most difficult task. The captive screws could wobble by several millimeters, while the screwdriver’s weight and the pull of its cable created recoil. A one-millimeter movement of the hand could result in an error of three to four millimeters at the tool tip.
The team also introduced disturbances during testing, including pushes to the robotic arm, shifts in the fixture base, and changes in lighting. The hand combined visual and tactile feedback to correct its movements in real time. If a screw failed to engage, it could withdraw and realign automatically.
Reinforcement learning, a training method based on trial and feedback, helped improve individual operations on the physical robot. Meanwhile, a software agent framework designed for industrial deployment coordinated the full sequence, checking each step, reversing failed operations, and tracking progress across workstations.
Agilink CEO Qiao Tianjie said decades of refinement have enabled conventional grippers to handle 70–80% of predictable industrial operations. Dexterous hands are intended to tackle the remaining 20–30% of complex, precision-intensive tasks that conventional grippers struggle with. A general-purpose dexterous hand could serve multiple processes, reducing equipment retrofitting costs and the limitations of machines designed for a single task.
The trial illustrated how the hardware and software framework could work together at a factory station. It also highlighted a persistent challenge facing the industry: the shortage of suitable training data.
Agilink’s chief scientist for embodied intelligence, Huang Bidan, said most dexterous manipulation models rely on visual datasets that lack physical interaction information, such as force feedback and tactile signals indicating slippage. Models that perform well in simulation can fail when deployed in the physical world. Factors such as temperature, shear force, and deformation at points of contact are difficult to simulate and need to be captured through real-world interactions.
The two main approaches to data collection have different limitations. Teleoperating a physical robot produces high-quality data but is expensive, requires skilled operators, and is difficult to scale. First-person recordings of human hands are easier to collect across different settings, but visual tracking alone offers limited precision and cannot capture tactile feedback. Such recordings are therefore more useful for pretraining than for teaching precise manipulation.
At the event, Agilink launched its OmniGlove Pro data collection glove and began volume deliveries. The company aims to evaluate data quality based on whether a robot can use the collected information to complete a task, rather than how closely recorded movements resemble those of a human hand.
The glove combines electromagnetic field (EMF) tracking of fingertip positions with data from inertial measurement units (IMUs) to reduce interference. Densely arranged tactile sensors cover the hand, while the system captures hand position and touch signals at high frame rates. The glove is lightweight and made of fabric. EMF tracking provides a reference for fingertip positions, while tactile sensors help compensate for positioning errors caused by changes in the fabric’s shape.
The fabric design is intended to interfere less with everyday activities and support prolonged, large-scale data collection. OmniHand robotic hands also come equipped with fabric tactile gloves. Matching the shape and sensor placement of the human and robotic gloves is intended to reduce information loss when translating human demonstrations into robotic actions.
Agilink said cumulative production of its OmniHand series exceeded 15,000 units as of October. Qiao considers production volumes in the tens of thousands a prerequisite for broader commercial deployment, arguing that such scale is necessary to stabilize supply chains and manufacturing yields while bringing down costs.
The company is developing its hands through industrial and commercial service deployments. Its robotic hands are already used regularly for commercial interactions at Chimelong in Zhuhai, while the dexterous hand station at Wolong Electric Drive is expected to enter regular production operations by the end of 2026.
Agilink aims to use data collected from these deployments to improve its models and expand the range of tasks its hands can perform. At the event, the company also announced plans to make its tactile data, manipulation models, and industrial deployment framework available to robot manufacturers, system integrators, and model developers, with the aim of expanding data collection and practical applications.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Huang Nan for 36Kr.