Robots entering factories is the clearest main line of the embodied-AI industry this year. A factory owner, asked how robots perform, said that for an export business with fast-changing orders, the line must shift constantly. Robots handle fixed tasks, but change the product and much has to be rebuilt.
That is the real problem after robots enter the factory: most still do one class of job in one set scenario. Change the task, from box moving to small-part feeding, and the gripper may need to become a dexterous hand, with development redone.
LeJu hit these issues repeatedly while deploying robots in factories, so it and Heiman released an industrial solution built around a general humanoid body and a general dexterous hand.
Doing several jobs is not the same as being general
Humanoids in factories now do single-point tasks: handling, sorting, loading and unloading. But those abilities depend on specific bodies and end-effectors. Change the material and conflict appears fast.
Small-part feeding needs precise pinching. Bin-picking or large-item handling needs high load and stable grip. Precise operation and heavy load pull the hand in opposite directions. A gripper carries weight but lacks finesse. A high-DOF hand is flexible but load-limited. Change the task and engineers may re-select, redesign and re-machine, cycles counted in weeks or months.
The humanoid’s value is adapting to human-built environments and taking different tasks across stations. If the end-effector stays highly custom, generality stalls at the last few centimetres. To let one robot keep doing different jobs, it must clear three gates: hardware compatible with different materials, models that reuse learned ability, and toolchains that shorten new-task deployment.
The dexterous hand’s bar is not more DOF
LeJu and Heiman’s answer is one hardware set covering both fine and heavy tasks. The hand uses a variable-configuration four-finger design, switching among 7-plus-N grasp forms, two-finger pinch, internal support, enveloping grip, two-hand carry. Single-hand load at least 24 kg, repeat positioning precision within 0.03 mm.
Notably the hand did not pile on parameters: 11 fully active DOF is enough. LeJu’s reason is direct. More DOF means more control difficulty and compute. The real bar for a dexterous hand is not how many joints, but whether you can control them.
Four fingers is an engineering trade too. Five-finger grasping often leaves one finger unloaded, redundant. Three fingers deflect when gripping slender objects like pencils. Four covers three-finger function while cutting five-finger redundancy, balancing grip stability and control complexity.
At the end-effector layer, a custom gesture editor inserts keyframes and adjusts whole trajectories, switching between position and force control. Validated motions enter an industrial gesture library. Similar tasks call them and tweak on site. LeJu says this doubles data-collection efficiency, and some new-scene end-effector adaptation drops from weeks or months to hours.
From one general hand to a whole system
Controlling the hand alone is not generality. When a person grasps, moves or adjusts an object, arm, shoulder and waist cooperate. The robot needs hand, arm, waist and whole-body joints in concert.
Over the past two years LeJu’s Kuafu has entered real industrial scenes: bin handling at FAW Hongqi, continuous depalletising at Haichen Logistics, small-part feeding in tight stations at Zhaofeng.
Each real scene leaves reusable real-machine data. KUAVO-VLA trains on 600-plus hours of homogeneous high-quality real-machine data, covering 100 industrial tasks, pre-loading common abilities into a vertical model.
For new tasks the model need not relearn how to control Kuafu. Developers focus on material, station and process differences. In LeJu’s 25-task evaluation, KUAVO-VLA ranked first on both overall success rate, 48.27 per cent, and process score, 74.51 per cent, against pi0.5 and GR00T N1.7.
LETools catches the next baton: data processing, model training and one-click deployment. The point is not one clever hand but a system where hardware, data, model and toolchain compound, so the next new task costs hours, not months.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience.
Translated and adapted from OFweek (https://robot.ofweek.com/2026-09/ART-8321202-8110-30701871.html).