WUJI Hand 2 用仿真训练实现真机转笔
Another beautiful robotic hand.
WUJI 训练其 20 关节机器人手 WUJI Hand 2 在物理模拟器中转笔，再把同一策略部署到真机，并在 IROS 2026 前后展示了配套的传感器手套。训练基于 MuJoCo 上的 mjlab 工具包，用 PPO 同时运行 4096 个虚拟手，学习 5 组共 40 个转笔动作；真机上策略以每秒 50 次向 20 个关节发送目标，工业相机追踪 250mm 3D 打印笔两端标记与腕部标签。WUJI 以 Apache 2.0 开源了代码、动作片段、训练模型和可打印笔文件。
查看原帖Pen spinning on WUJI Hand 2, from simulation training to real-world execution. We use motion-reference tracking and reinforcement learning to bring a simulation-trained policy to the physical hand. Code and setup guides are open source for developers and researchers to build on. Code: https://t.co/EVDLueqNVS
正文
当前显示原文，中文全文译文尚未就绪。
Another beautiful robotic hand.
WUJI trained its 20-joint robot hand, WUJI Hand 2, to spin a pen inside a physics simulator, then ran that same trained policy on the real hand.
It showed this around IROS 2026 (the International Conference on Intelligent Robots and Systems), together with a sensor glove that lets a person control the hand live.
Pen spinning is a hard test for a robot hand, because the pen keeps rolling, sliding and passing between fingers, and one finger moving a bit late drops it. @wuji_global’s Hand 2 has 20 joints, four per finger, each independently driven by its own motor, and supports a 1,000 Hz control rate.
Training used mjlab, a robot-learning toolkit built on the MuJoCo physics simulator, running 4,096 virtual copies of the hand at the same time with PPO (Proximal Policy Optimization), a reinforcement learning method where the AI improves by trial and error and gets rewarded for good moves.
The AI learns to follow 40 recorded pen motions in 5 groups, from a simple turn around 1 axis up to continuous spins across several axes, and it keeps correcting its fingers whenever the pen drifts from the plan.
On the real hand, the policy sends targets to all 20 joints 50 times a second, while an industrial camera tracks printed markers on both ends of a 250mm 3D-printed pen plus a tag on the wrist, so the AI always knows where the pen is.
The glove covers the other way of teaching a robot: 5 electromagnetic sensors at the fingertips track each finger's position and angle 120 times a second with about 10ms delay or less, and software turns that into joint angles the robot hand copies.
A 526-point pressure grid on the glove's palm also records how hard the person grips, which makes it useful for collecting training data for robots that learn by copying humans.
WUJI released the code, motion clips, trained models and printable pen files under the open Apache 2.0 license, so a lab with a WUJI Hand 2, a camera and an NVIDIA GPU can repeat the whole setup.