AgiBot's OmniHand Wins 7 Golds, But Dexterous Hands Face Factory Test
At the second World Humanoid Robot Games, OmniHand, a dexterous hand from AgiBot subsidiary Critical Point, won seven golds in fine manipulation events such as powder weighing, with error within 0.5 grams, and bottle opening and cap prying. Compared with bipedal walking, running and somersaults, these seemingly minor actions drew more excitement.
Qiao Tianjie, CEO of Shanghai Critical Point Innovation Intelligent Technology Co., Ltd., told Xinmin Evening News that for robots to truly enter factories, malls and daily life, they need not only vision and intelligence but also hands that can use human tools and complete real work.
Why is the dexterous hand the crux? In past industrial automation, robot arms mostly used pneumatic two-finger grippers or electromagnetic suction cups. These can only open and close, acting as highly specialized tools. In the embodied AI era, humanoid robots are expected to enter thousands of industries and even homes, making dexterous hands unavoidable.
Qiao said the human physical world—from door handles and steering wheels to screwdrivers and cups—is designed around human hands. For robots to use human tools, their hands must closely resemble human hands. More importantly, in AI system architecture, dexterous hands are not only output devices for work but also critical input devices for perception. How much friction does a glass cup surface have? What is an apple's skin stiffness? What is the resistance feedback halfway through tightening a screw? Such rich physical information cannot be obtained by vision alone; it must be sensed in real time at the moment of contact and sent back to the brain.
Trade-offs in degrees of freedom. Human hands have about 27 degrees of freedom. Replicating this dexterity in machines challenges manufacturing, algorithms and sensing. OmniHand, which won gold, achieved 16 degrees of freedom in a compact space. Qiao said simply pursuing high DOF can lead to a technical misconception. Current dexterous hands roughly fall into low, medium and high tiers: 6-DOF low-end hands often pose stiffly when holding nails or spoons and cannot even hold tools steadily; high-end hands with over 20 active DOF can perform highly realistic motions in labs, but to pack more parts into tight spaces their mechanical design approaches limits. At the competition, some high-DOF hands suddenly failed or jammed due to motor overheating. Once DOF is too high, reliability and stability inevitably suffer.
Where are the bottlenecks? First is the dimensional explosion and generalization challenge of algorithm models. Qiao explained that one dexterous hand has a dozen to twenty DOF; two hands together approach 40 DOF, roughly equal to all joints in the body excluding hands. This exponential jump in complexity poses huge challenges for model training. Second is the global industrial chain and underlying technology contest. In hardware maturity, domestic dexterous hands lead foreign counterparts, benefiting from China's booming new energy vehicle industry and its mature supply chain. But in model algorithms and top AI talent, technical bottlenecks remain. Third is the commercialization gap from lab to market. In real industrial scenarios, customers strictly calculate ROI, cycle time and mean time between failures. If a dexterous hand crashes and stops a customer's production line, the loss is enormous.
The first large-scale users of dexterous hands may not be humanoid robots. Qiao said there is a counterintuitive industry consensus: the first large-scale application of dexterous hands may not be bipedal humanoid robots but combinations of robot arms and dexterous hands. For many business models to run healthily, cost control is central. Most logistics sorting and industrial assembly scenarios are fixed at a workstation. There, an arm plus hand is the best working system, and buying a full humanoid robot is unnecessary. Traditional grippers can handle 70% to 80% of simple pick-and-place tasks, but complex processes such as using screwdrivers, drills, wire harness plugging and fine cable routing still rely on manual labor. As mid-tier dexterous hands become more cost-effective, these long-standing automation blind spots will be cleared one by one. Critical Point's dexterous hands have reached mass production of over 10,000 units, supplying AgiBot's own robots and serving industrial and commercial customers in vertical fields at home and abroad.