The DexGEM platform centers on a technical closed loop for physical AI, detailed in a research paper titled "Dexterous Robot Manipulation from Human Demonstrations via Contact-Anchored Retargeting and Residual Policy Learning." By reconstructing human movements from motion capture and first-person video within a simulation environment, the system transfers complex skills—such as two-hand coordination and continuous-contact manipulation—to physical hardware. This approach bypasses traditional teleoperation, allowing robots to learn nuanced dexterity independently.

Supporting this software milestone is the DexHand021 Pro, a 22-degree-of-freedom robotic hand designed for research-grade durability and sensing. The device integrates an actively driven 2-DoF wrist with a dual-tendon bidirectional drive, facilitating precise force control and passive compliance. To manage interaction, the hand features a 3D tactile envelope array capable of 0.01 N normal-force resolution at 200 Hz, alongside a 160-degree wide-angle palm camera. An onboard edge AI chip processes this multimodal data to execute millisecond-level decisions, a capability validated by over 300,000 durability test cycles.