x2robot launches TwinDex with zero real-robot teleop data

The launch comes months after Nvidia robotics head Jim Fan said in a public speech in May that vision-language-action models and teleoperation are dead. At the time, the remark was widely seen as another AI hype statement, but the TwinDex demo described by QbitAI is an early sign that the claim is beginning to have practical substance.

In a continuous one-take demo, TwinDex opened bottles, sampled liquids with a pipette, handled test tubes, used a glass rod for pouring and swirled a flask to observe the result. The routine comprises 24 sub-actions spanning three kinds of tools, with repeated two-hand coordination and tool changes. The company says each step requires millimeter-level positioning and stable force control.

The hand's demonstrations highlight capabilities that have long been difficult for robotic hands. To open a toolbox, two index fingers must enter narrow slots, press down to release the latch, then pinch a handle. In syringe use, two fingers fix the barrel while the thumb pushes the plunger and controls direction and force. A third finger provides extra contact for supporting long tools, such as a broom, and in page turning the thumb first separates the top sheet before the other hand takes over.

TwinDex is not a standalone gripper. It combines a wearable three-finger exoskeleton for data acquisition, a robot end effector, a data-processing pipeline and a training recipe. The acquisition and execution hardware share the same kinematic structure, so captured actions can be mapped directly to the robot without converting between unrelated devices.

During collection, operators wear the exoskeleton and interact with real objects, receiving natural force feedback. No physical robot needs to be occupied, and multiple operators can collect data in different locations. The system records vision, joint states and wrist pose, then aligns the records through calibration and time synchronization before training.

x2robot says the system produces about 5.3 times as many usable trajectories per unit of time as traditional real-robot teleoperation. In experiments described by the company, policies trained with non-embodied data and policies trained with real-robot teleoperation data improved at the same rate as data volume increased and converged to comparable levels. In the tested tasks, non-embodied data could almost completely replace real-robot teleoperation samples.

The company explains the design through three features: dexterity, consistency and scalability. After comparing candidate designs on basic grasping, in-place twisting, tool use and in-palm manipulation, it chose a three-finger, nine-degree-of-freedom layout as a balance between the simplicity of a two-finger gripper and the complexity of a five-finger humanoid hand. The extra contact point from the third finger is especially valuable for stable tool handling. Consistency is the central idea: instead of aligning collected data to the robot after the fact, x2robot aligned the acquisition device and the robot manipulator in motion, contact, visual observation, accuracy and timing before data generation. Scalability follows because data collection no longer requires occupying a robot for every demonstration.

The report cautions that teleoperation is not dead yet. Real-robot data remains important for broad task coverage and data diversity, and the data recipes for embodied AI are still unsettled. But the strong link between high-quality robot data and physical teleoperation is loosening. x2robot's answer, the report says, is that the era of starting every new task with real-robot teleoperation data is ending.