Zibianliang's TwinDex completes fine chemistry lab tasks with zero teleoperation data

On September 3, Chinese robotics company Zibianliang (自变量, project hub x2robot.com) unveiled TwinDex, a new dexterous operating system. The centerpiece is a continuous, single-shot chemistry experiment: with zero real-robot teleoperation data in post-training, the system performed fine manipulations such as unscrewing caps and pushing syringe plungers using only a few hundred bodyless demonstrations. The run spans 24 sub-actions across three tools, with repeated bimanual coordination and tool switches, each step demanding millimeter-level positioning and stable force control.

QbitAI's report places the launch inside a larger debate. In May, Nvidia's robotics lead Jim Fan declared at a public talk that VLA models and teleoperation are dead, a claim most people dismissed as another hot take. A few months on, TwinDex completing fine operations without real-robot teleop data is read as early evidence that the argument is coming true.

The demo showcases three historically hard categories of dexterity. Precision fine manipulation: both index fingers align to narrow latches on a toolbox, insert and flick them open, then pinch the handle to lift the lid, while syringe work sees two fingers brace the barrel and the thumb drive the plunger. Three-finger collaboration: in a broom-and-dustpan task, three fingers wrap the handle in an envelope, sharing support and control. And human-like flexible in-hand adjustment: caps are twisted by pinching with thumb and index finger and rotating via finger-side swing rather than large wrist motion, while page-turning begins with the thumb rubbing the top page loose before pinching, handoff and placement.

The piece notes that such fine operations previously conjured images of complex five-finger hands and vast teleop datasets, but TwinDex shows three fingers are enough for many tasks. The three-finger, nine-degree-of-freedom layout is a deliberate compromise among dexterity, stability and engineering complexity: two-finger grippers struggle with twisting and in-hand manipulation, while five-finger designs multiply joints, actuators, sensors and calibration overhead. After meta-operation tests covering grasping, in-place twisting, tool use and in-hand work against several candidate configurations, Zibianliang settled on three fingers and nine DOF as the sweet spot.

TwinDex is not only an end-effector; it doubles as a wearable bodyless data-capture system. Unlike setups that use one set of hardware for collection and another for execution, TwinDex keeps capture and execution isomorphic: the same three fingers that work also collect data, a what-you-capture-is-what-you-get design. Because motion patterns, contact modes, visual observation and time synchronization match on both sides, operator actions map directly onto the robot, effectively moving embodiment alignment from the data-processing stage to before data is even generated.

The efficiency numbers are striking: TwinDex yields about 5.3x the usable trajectories per unit of time compared with conventional real-robot teleoperation. The wearable exoskeleton-style rig frees collection from physical robots, so multiple operators can record simultaneously in different locations, and the pipeline logs multimodal signals including vision, joint states and wrist poses, unifying them through calibration, time synchronization and normalization into a single training space. Model architecture and training are also designed to tolerate capture noise and drift.

More important is data utility. Zibianliang found that as dataset size grows, policies trained on bodyless data and on real teleop data improve at the same rate and converge to similar performance; in these experiments, bodyless data could substitute for nearly 100% of real teleop data in training. The long-held assumption that every new task must start from real-robot teleoperation demonstrations is being broken.

QbitAI's verdict is that TwinDex matters not just because it captures data faster, but because it demonstrates that the highest-quality data closest to real robot motion does not have to be produced by a real robot. To be sure, embodied-data recipes are far from converged, task coverage and diversity remain limited, and nobody can yet declare teleoperation dead. Still, one thing is increasingly clear. The strong coupling between high-quality data and real-robot teleoperation is loosening. The open questions are whether TwinDex reproduces these results across more tasks, and whether bodyless data can develop a scaling law of its own once volumes grow.