ZiBian Unveils TwinDEX Dexterous Hand System for Embodied Intelligence Without Proprioceptive Data

Recently, Zibian Robotics unveiled TwinDEX—a pair of twin three-fingered, nine-degree-of-freedom (DoF) dexterous manipulation hands—comprising a wearable version for data collection and a robot-mounted version for real-world deployment. This system marks the first achievement of purely embodiment-free, zero real-robot teleoperation data-driven dexterous manipulation. It enables foundation models to accomplish complex tasks after fine-tuning with only hundreds of embodiment-free data samples, without using any real-robot teleoperation data throughout the entire process. TwinDEX is designed around three core principles: dexterity, consistency, and scalability. Its three-finger, nine-DoF architecture features seven active DoFs, striking a balance among dexterity, reliability, and cost. According to Zibian, the three fingers—thumb, index, and middle—work collaboratively to cover most manipulation tasks encountered in everyday scenarios, forming the "minimum viable solution" for dexterous manipulation. In terms of consistency, TwinDEX maintains uniformity across multiple dimensions: kinematically, it preserves identical DoFs, joint axes, and link proportions; in contact mechanics, it standardizes contact materials, geometry, and surface properties while integrating corresponding tactile sensors; visually, it ensures consistent appearance; and it optimizes critical precision metrics such as joint accuracy, wrist positioning, jitter, and drift. Anchored by target application performance, the system reverse-engineers the required consistency specifications for the data acquisition side, effectively addressing common issues like kinematic mismatches and precision loss in cross-embodiment data collection. Regarding scalability, TwinDEX employs a wearable three-finger exoskeleton as its data acquisition interface, decoupling data collection from robot hardware and fixed environments. Reportedly, “one operator, one table, and one exoskeleton” constitute a complete data collection unit. The exoskeleton provides intuitive force feedback, enabling natural and precise contact interactions, while its modular design allows parallel scaling simply by adding more devices and operators. In relevant tasks, TwinDEX achieves a data collection efficiency 5.3 times higher per hour than traditional real-robot teleoperation. Zibian stated that its next step is to expand data scale—from single desktop scenarios to open environments, and from fixed settings to long-tail distributions—thereby providing richer and more diverse data sources for robotic model training.