X Square Robot Unveils TwinDEX for Scalable Dexterous Manipulation

Dexterous manipulation requires more than simple pick-and-place movements; it demands the ability to twist, pour, and adjust through continuous contact. Traditionally, gathering training data has been a bottleneck, as on-robot teleoperation remains expensive, slow, and restricted to specific workspaces. While robot-free data collection is easier to scale, it often introduces an embodiment gap where kinematic differences between the human hand and the robot degrade the quality of learned policies.

TwinDEX resolves this by aligning the collection and deployment hardware through a shared three-finger, nine-degree-of-freedom architecture. The system synchronizes multi-view RGB inputs, wrist poses, and tactile signals to ensure that data captured by a human operator maps directly to the robot's joint space without requiring complex retargeting. This allows operators to collect data in diverse environments like offices or kitchens without needing a robot present.

In testing, TwinDEX achieved up to 5.3 times the throughput of conventional on-robot teleoperation. The system successfully executed a complex, multi-stage chemistry experiment—involving tool switching and liquid transfer—using only a few hundred robot-free episodes. By removing the dependency on expensive robotic hardware for data collection, the framework allows for the rapid, large-scale training of models capable of long-horizon, fine-grained tasks.