X Square Robot Launches TwinDEX for Robot-Free Dexterous Manipulation Training

In contact-rich tasks such as twisting, inserting, pouring, opening latches, and using tools, robot actions depend not only on hand position but also on details like finger contact forces and continuous corrections. X Square Robot points out that data collection for such tasks faces a dilemma: demonstrations from robot teleoperation match the target hardware but are expensive to scale because they occupy robots, require prepared workspaces, and may depend on skilled operators; robot-free data collection is easy to scale, but differences between the collection device and the robot can make demonstrations difficult to transfer.

The core idea of TwinDEX is to make the wearable collection device and the robot end-effector consistent across multiple features that affect policy training. According to the company, both share kinematic chains, joint axes, link proportions, contact geometry, surface materials, visual appearance, and sensor layout, allowing finger states measured by the wearable to be directly mapped to the robot's joint space without complex hand-to-robot retargeting.

The system also synchronizes camera input, wrist pose, finger states, and tactile signals, and measures latency between sensing, model inference, and execution, making the timing between observations and actions during training closer to actual deployment.

During data collection, operators wear the exoskeleton and interact directly with real objects, without robot involvement. A single collection unit requires only one operator, a table, and the wearable system, so multiple operators can collect demonstration data in parallel at locations like offices, kitchens, and workstations. This mode aims to address the scalability issues of teleoperation, but the trade-off is that differences between the collection device and the robot can affect transfer, and TwinDEX attempts to eliminate this barrier through hardware matching.

X Square Robot reports that TwinDEX achieves up to five times the effective data collection throughput of robot teleoperation. This figure comes from the company's own testing and is a source claim, not yet independently verified by third parties.

The company evaluated TwinDEX on various tasks, including unscrewing bottle caps, sweeping with a broom and dustpan, removing and opening a book, releasing toolbox latches, and operating a syringe. Additionally, TwinDEX was tested in a standardized chemistry experiment with 24 sub-actions, where the robot handled containers, spoons, pipettes, liquids, solids, and glass rods, completing the task autonomously in a single continuous run while switching tools and coordinating both hands.

Currently, information about TwinDEX primarily comes from company releases and media reports, lacking third-party evaluations or peer-reviewed data. The 5.3x throughput improvement is a self-reported result, and its test conditions, comparison baseline, and statistical significance have not been disclosed. Furthermore, the system's long-term stability and generalization in real-world complex environments have not been verified, and the success rate of the chemistry experiment has not been disclosed.

While the hardware matching strategy theoretically simplifies data transfer, its actual effectiveness depends on factors such as sensor accuracy, timing synchronization, and operator skill, details of which are not elaborated in available materials.