Building generalist robots for real-world deployment
The Hook
“Twisting open a bottle cap, operating a pipette, and a scooper, coordinating both hands. This pair of robot hands is autonomously performing a chemistry experiment. Not a single piece of its training data came from a physical robot.”
Problem Agitation
“Behind this lies a long-standing dilemma in dexterous manipulation: collecting data through on-robot teleoperation is costly and hard to scale; collecting large volumes of data through low-cost alternatives yields data that is difficult to transfer directly to the target hardware.”
Product Reveal
“Today we introduce TwinDEX: a pair of co-designed three-finger, nine-DoF dexterous manipulation interface. One wearable for data collection, one for robot deployment. Together with an integrated data processing pipeline and policy training workflow, they form a complete collection-to-deployment system.”
Feature Teaser
“The end-effector design is the result of a careful engineering trade-off. Additional fingers raise the ceiling of dexterity, but impose escalating demands on spatial packing, actuator torque density, and deployment reliability. By evaluating a broad set of typical basic human manipulation primitives, we determined that a three-finger, nine-DoF design, seven of which are actively driven, offers the best balance.”
Feature Teaser
“Consistency quantifies the alignment between collection and deployment devices across kinematics, contact dynamics, and visual observation. It directly determines whether robot-free data can substitute for teleoperated data in policy learning. We anchor on closed-loop deployment performance and work backward to derive quantitative requirements.”
Feature Teaser
“The collection device is a wearable three-finger exoskeleton that decouples data collection from both a robot embodiment and a fixed workspace. One operator, one table, and one exoskeleton constitute a complete collection unit. The wearable form factor further provides the operator with intuitive force feedback, enabling high-quality collection of contact-rich tasks.”
Product Reveal
“Now, back to the chemistry experiment. Twenty-four sub-actions—twisting open a bottle cap, operating a scooper and pipette, guiding liquid flow with a stirring rod, bimanual coordination across diverse fine objects, each demanding millimeter-level precision and stable force control. All data is collected from scratch through the wearable device alone, just a few hundred episodes, with zero on-robot data.”
Call to Action
“On a standardized benchmark covering representative manipulation primitives, TwinDEX achieves a 5.3 times improvement in collection efficiency of on-robot teleoperation while maintaining comparable data quality. TwinDEX was designed for scalable data collection from day one. As collection scales up, the system continues to support tasks of increasing complexity, making data scaling a viable path toward general-purpose dexterous manipulation.”
Synthesizable in Remotion · key components of the twindex-robotics-showcase template.
Dual-frame comparison running simultaneous teleoperation and autonomous playback.
Wrap two Video elements inside side-by-side AbsoluteFill containers with matching trim frames and rounded inner borders.
Large numerical HUD counter incrementing seamlessly as task segments complete.
Interpolate current frame to integer index mapped to an array of timestamps, rendering bold monospace typography.
Glowing SVG polyline graph drawing on screen with semi-transparent confidence corridors.
Use strokeDashoffset interpolation mapped to spring-driven frame values over an SVG path with linear gradient fills.
Purpose. Demonstrate high autonomy in laboratory conditions without physical teleop training.
Execution. High-precision robotic fingers handling delicate glassware at 3X speed.
Purpose. Articulate the scaling and transfer trade-offs of existing methods.
Execution. Split-view archival footage of teleoperation rigs and handheld collectors.
Purpose. Present TwinDEX through Dexterity, Consistency, and Scalability.
Execution. Mirror-action split view followed by detailed metric diagrams.
Purpose. Validate robustness across a continuous 24-step manipulation chain.
Execution. 8X speed continuous run paired with an assertive on-screen subaction index.
Purpose. Establish vision of scalable generalized manipulation data.
Execution. Scaling mosaic grid dissolving into the corporate branding identity.