Author | Zhou Jiaqi Editor | Moying
Robotic hands are becoming increasingly dexterous. Grasping objects, tightening screws, operating tools... An increasing number of actions that previously required human hands can now be performed by dexterous hands. However, the situation changes when they truly enter factories. Currently, industrial automation still relies mainly on grippers for specific processes, and applications of dexterous hands are mostly concentrated on single operations. Enabling a single dexterous hand to adapt to different tools and continuously complete an entire workflow remains a challenge.
Zhidx reported on October 29 that today, at the AGIBOT Day Dexterous Manipulation Sub-forum, AGILINK live-linked with the Wolong Hill factory to showcase its latest industrial POC results: Through the combination of "robotic arm + dexterous hand," it continuously completed a dexterous hand performance test process consisting of 72 steps. This process involves various operations such as loading/unloading, scanning codes, grasping screwdrivers, tightening/loosening screws, and pressing buttons.
Phase one validation data shows that the task success rate reached 98%, with a single operation taking only 8.5 minutes, which is already close to the human working cycle of approximately8 minutes, and it can run continuously for 8–12 hours. Meanwhile, AGILINK also launched the new teleoperation and data collection integrated glove OmniGlove Pro, capable of synchronously collecting human hand movements and tactile information, and has completed Replay verification of operational data on actual robots. The unveiling of this industrial POC and new data collection product further demonstrates AGILINK's full-chain technology layout centered around dexterous hand hardware, data, models, and actual deployment.
01.72 Continuous Steps, Work Efficiency Close to Humans
At the dexterous hand performance testing station in Wolong Hill, one dexterous hand is testing another. Upon closer inspection, the division of labor between the two hands is completely different. The hand mounted on the end of the robotic arm is responsible for completing the entire operation; waiting on the conveyor belt for inspection are the dexterous hand products about to leave the factory from AGILINK. Wu Mingdong, Algorithm Director at AGILINK, stated that AGILINK is using its own dexterous hands to test its own dexterous hands. The robotic arm drives the end-effector dexterous hand to grasp the "hand" under test, places it on the testing equipment, and then scans the code to identify information.
Immediately after, the dexterous hand picks up the adjacent screwdriver, aligns it with the screw hole, and fixes the product under test onto the testing station. After returning the tool, the robot presses the start button to begin performance testing. Once the test is complete, the dexterous hand presses the stop button, picks up the screwdriver again, loosens the fixing screws, and returns the product under test to the conveyor belt.
This process involves three different types of operations: flexible assembly of the dexterous hand under test, tightening/loosening screws with a screwdriver, and operating the dashboard of the testing equipment. These respectively test capabilities such as complex object grasping and positioning, fine force control and tool operation, and button interaction. The entire process contains 72 operational steps, all autonomously completed by one robotic arm paired with one dexterous hand.
According to on-site introductions, AGILINK previously attempted to use grippers, but the reaction force during screw tightening easily caused the screwdriver to slip from the grip. If the tool were fixed to the end of the robotic arm, it could not兼顾 tasks such as grasping products or operating dashboards. Therefore, AGILINK chose to use a five-finger dexterous hand, capable of continuously completing different types of operations without changing the end-effector.
Wu Mingdong revealed that the task success rate for the Phase One POC at the station reached 98%, and the robot can operate continuously without interruption for up to 12 hours. Moreover, the robot completes the entire 72-step test process fully autonomously in only 8.5 minutes, while humans take approximately 8 minutes to complete the same work.
The difference is only 30 seconds, meaning the robot's single operation time is only about 6.25% higher than that of a human. This implies that a robotic arm paired with a dexterous hand can already continuously perform tasks such as grasping, clamping, screw tightening/loosening, and equipment operation at a working pace close to that of humans. Compared to deploying a full humanoid robot directly, adopting the "robotic arm + dexterous hand" combination also helps control equipment investment costs. It is understood that AGILINK plans to officially integrate this solution into production lines within the year to further verify its long-term stability and deployment cost in real production. Regarding this application scenario, Qiao Tianjie, Vice President of AGIBOT and CEO of AGILINK, stated that a pair of hands assembling the next pair of hands on the production line is the future of AGILINK.
02.Sub-Millimeter Precision Alignment, Autonomous Recovery Under Interference
These seemingly ordinary and easy operations pose significant technical challenges when actually executed by robots. First is the precise alignment of the screwdriver. In actual operations, the screwdriver bit needs to accurately enter the screw head, requiring sub-millimeter alignment precision. Even experienced workers need to constantly adjust the position and angle of the tool to ensure accurate alignment.
To allow for repeated testing, the station uses reusable anti-loosening screws, which cause millimeter-level wobble in the holes, further increasing the difficulty of alignment. At the same time, the screwdriver itself is large and heavy, with connecting cables causing drag, and there is recoil force during tightening. Even if the posture of the dexterous hand holding the screwdriver deviates by just 1 millimeter, it may cause the gun head to offset by 3–4 millimeters.
In this industrial POC, AGILINK's dexterous hand was able to autonomously complete this delicate operation. On site, AGILINK also demonstrated the robot's autonomous correction and failure recovery capabilities by artificially creating interference.
Staff pushed the robotic arm while the robot was tightening screws, causing the screwdriver to deviate from its original position. It can be seen that the robot did not stop working but autonomously adjusted its position and posture, re-found the screw hole, and continued to complete the tightening operation.
In another test, staff artificially adjusted the fixture base, causing the screw hole position to shift, resulting in the screw being only partially tightened. The robot identified the operation failure through tightening force and visual information, then re-aligned the hole position, tightened the screw again, and finally completed the task. To handle these complex situations, AGILINK adopted a solution combining real-machine reinforcement learning with an agent system. Specifically, reinforcement learning allows the robot to adjust actions in real-time based on visual and tactile feedback, improving precise alignment and anti-interference capabilities.
AGILINK disclosed that after training via real-machine reinforcement learning, the robot's screw-tightening speed is more than 4 times faster than teleoperation, and alignment movements are more direct. For operation failures that have occurred, AGILINK handles them through the industrial deployment-state agent framework Omni-Maestro. This system can combine visual, torque, and station equipment feedback to determine whether the task is complete, and upon anomaly detection, re-execute or backtrack to the error step for repair.
03.Complementing High-Quality Data, Accelerating Scaled "Deployment State"
To enable dexterous manipulation to move toward more industrial scenarios and drive the true landing of the "deployment state," obtaining high-quality operational data is a key link.
Currently, Ego data is an important path for scaled training in embodied intelligence. Compared to relying on data collection from actual robots, obtaining data by recording human daily operations makes it easier to scale data volume. However, traditional Ego data has limitations. These solutions primarily rely on cameras to record human hand movements and estimate hand posture via algorithms, making them susceptible to factors like visual occlusion. Additionally, due to the lack of tactile information, it is difficult to restore the state of contact between the human hand and objects.
In response to this, the AGILINK team independently developed the data collection glove OmniGlove Pro and simultaneously initiated mass production delivery.
This glove can simultaneously collect human hand position, posture, and tactile information, packaging them uniformly for output to provide data support for subsequent robot operations. Specifically, regarding motion collection, each fingertip of the OmniGlove Pro features EMF magnetic field positioning and IMU anti-interference fusion technology, achieving a positioning distance accuracy of 0.1mm! Regarding tactile collection, this glove is configured with 746 tactile sensing points, with a spatial resolution reaching 3mm, allowing for more accurate recording of contact between the hand and objects.
After processing and cross-embodiment action retargeting, this data can be Replayed on actual robots to complete tasks identical to human demonstrations.
Huang Bidan, Chief Embodied Scientist at AGILINK, refers to this type of data as "Real-Machine Ready" Ego data.
This data combines the scaling advantages of Ego data, the high quality of real-machine data, and the authentic feel of human operations. With the integration of these three elements, we can expect to see accelerated implementation of scaled deployment for dexterous manipulation.
04.Conclusion: Dexterous Hands Moving Into Factories, Speaking Through Actual Performance
The release of this industrial POC and OmniGlove Pro corresponds to two key links in AGILINK's implementation of dexterous manipulation: the former verifies the ability of dexterous hands to continuously complete complex tasks through real stations; the latter enables human operational data to be used more efficiently on actual robots, further bridging data collection and actual operation.
As dexterous hands gradually enter industrial scenarios, maintaining operational precision under complex conditions, stabilizing work at a pace close to humans, and replicating these capabilities across more scenarios will become key to their scaled application.