What if a dexterous robot could learn a 20+ step chemistry experiment without a single on-robot training demo? Meet TwinDEX: a pair of co-designed, three-finger, nine-DoF dexterous manipulation interface: one wearable for data collection, one for robot deployment. The twinned design shares identical kinematics, contact surfaces, visual appearance, and... sensors across collection and deployment — keeping observations and actions aligned end to end. Trained from scratch on only a few hundred wearable demonstrations - with zero on-robot training or intervention data - TwinDEX completed a standardized chemistry experiment involving tool switches, fine force control, and bimanual coordination. Robot-free data showed comparable learning efficiency on the multi-task evaluation, TwinDEX delivered 5.3 times effective throughput than on-robot teleoperation. TwinDEX demonstrates that high-quality robot-free data can fully substitute for on-robot teleoperation data on challenging dexterous tasks — removing the dependency on real-robot hardware that has been the central bottleneck to scaling dexterous manipulation data. This was the proof-it phase. Now comes scale: what emerges at tens of thousands, or millions, of episodes? Watch the demo and read the technical blog: #TwinDEX #Robotics #EmbodiedAI #DexterousManipulationshow more

Read our full technical blog for details: TwinDEX reaches the strong capability-cost trade-off among the evaluated hand morphologies. Its 7-active-DoF, three-finger design preserves high dexterity across precision manipulation, in-hand manipulation, tool use, and human-machine interaction, while keeping the system reliable, wearable, calibratable, and scalable.

way more scalable than teleop

Honestly, this is awesome. No on-robot training data, then straight onto matched hardware! 👍👍👍

This robot is one step away from becoming Heisenberg 😁

you know this lil dude is knocking over some HCl

This looks really good. But how difficult is it to transfer data from a three-finger UMI setup to other embodiments? As tasks get more complex, could data collected with a three-finger UMI eventually become insufficient for more complex hands?

Great question. TwinDEX works because the wearable and robot hand are co-designed to match in kinematics, contact geometry, sensing, appearance, and timing. If you move to a very different hand, likely need retargeting, or a matched collector for that embodiment. Three fingers are our current sweet spot for daily dexterous tasks, not a universal limit for all future hands.

Thanks for the explanation, that was really helpful. UMI seems like one of the best approaches for robot data. As models improve at in-context learning, high-quality robot data collection systems like this will become even more valuable.

fantastic, another friction removed from robot training. and this looks very different from collecting generic human demonstrations and hoping a retargeting layer can reconstruct what the robot should have done.

The key result is that TwinDEX robot-free data follows the same data-efficiency trend as on-robot teleoperation data. As dataset size increases, policy performance scales at a comparable rate, showing that robot-free demonstrations preserve high learning value.

TwinDEX decouples data collection from the robot. Across long-horizon contact-rich tasks, one operator with one wearable system can collect demonstrations far more efficiently than on-robot teleoperation, while still producing deployment-aligned data.

TwinDEX provides a direct visual signal for correspondence quality: both hand URDF models are reprojected onto the head-camera image plane in real time from wrist poses and finger joint states. When the rendered hands stay aligned with the observed hands, it verifies kinematic calibration and cross-modal timing synchronization in the same closed loop used for data collection.

What follows is a full standardized chemistry experiment, from opening the reagent bottle and sampling, to guided pouring, mixing, swirling, and observation, executed autonomously by the policy in a single uncut take. All policies are trained from scratch using robot-free data only, with zero on-robot teleoperation data introduced at any stage.

Towards truly general-purpose dexterous manipulation. TwinDex！

Good job guys This is truly remarkable

Impressive work.

5.3x lines up with what we see. What got harder for us at volume was the people. Fifty contributors on the same setup, recordings that barely resemble each other.

Huge step for real-world AI.

Zero hardware bottleneck now

TwinDEX could make lab automation far more scalable

This could lower the barrier for training dexterous robots in many different settings

Robot-free training could remove a major bottleneck in dexterous manipulation.Scaling high-quality demonstrations without tying data collection to physical robots could accelerate how quickly complex skills are learned. 🤖🧠

Need to see how it scales to 10k episodes fr

Interesting approach to the data scaling problem. The real test will be how well this holds up across increasingly diverse, unstructured real-world tasks.

If robot-free demonstrations can deliver that kind of throughput without on-robot collection, scaling manipulation learning will be more practical.

The co-designed contact surfaces are what make the Sim2Real transfer actually work here. most teams try to remap visual tokens onto mismatched grippers and wonder why slipping happens.

bimanual coordination usually breaks zero-shot transfer, how clean was the data?

A wearable three-finger interface paired with a matching robot hand is a clever design choice.

High quality robot data is valuable, but collecting it efficiently matters too.

5.3x throughput without real robot hardware setup is crazy

Capturing precise finger-level behavior without needing the robot in the loop is the key idea here.

This is more convincing than just another short pick-and-place demo.

congrats to XSquare team 🔥

Thought zero on-robot training was hype until seeing identical kinematics across collection and deployment

Teleop pipelines almost all lose fidelity at the human→robot retarget step — you end up training on poses the robot sometimes can't physically hit. Matching the wearable's kinematics to the deployment hand doesn't shrink that error, it deletes the term. And reprojecting both hand URDFs onto the head-camera view as a live check is the same residual test photogrammetry uses for calibration — cheap, visual, catches drift before it poisons a session. Two good calls.

Three finger setup > 5 finger hype