PITTSBURGH, Sept. 30, 2026 /PRNewswire/ — DexRobot, a full-stack embodied AI company focused on dexterous manipulation, is showcasing its complete dexterous manipulation capabilities at IROS 2026 (Booth 507), including three accepted research papers; the flagship DexHand021 Pro with full-palm multimodal sensing; and DexGEM, a phased research milestone of its in-house models demonstrating a technical closed loop in dexterous manipulation physical AI.
DexGEM: A Research Milestone in Physical AI Dexterity
DexGEM is the platform behind one of three accepted papers: “Dexterous Robot Manipulation from Human Demonstrations via Contact-Anchored Retargeting and Residual Policy Learning.”
DexGEM learns dexterity from human hands. Demonstrations captured through motion capture and first-person video are reconstructed in simulation, adapted to robotic hands, and transferred to real robots, including exploration across different hand designs. Tasks run fully autonomously, with no teleoperation and no real-robot training data, spanning two-hand coordination such as object handovers and continuous-contact manipulation.
These results are built on DexRobot’s dexterous hardware (DexHand021 series), native perception, and manipulation datasets accumulated over years, trained through in-house models and validated in real interaction scenarios.
The DexGEM pipeline: converting human demonstrations into contact-rich robot policies, validated on real hardware with no real-robot training data
DexHand021 Pro: Research-Grade Dexterity for Real-World Tasks
DexHand021 Pro combines 22 DoF in a wrist-hand integrated architecture with an actively driven 2-DoF wrist, enabling natural wrist-hand coordination for turning, aligning, flipping, and tool use. Its dual-tendon bidirectional drive with active distal joints enables independent active flexion and extension with precise force control and passive compliance.
Full-domain multimodal perception covers palm, fingertips, and finger sides: a 3D tactile envelope array offers 0.01 N normal-force resolution at 200 Hz sampling, while a 5-megapixel, 160° wide-angle palm camera tracks contact, force, and slip to guide adjustments.
A dedicated edge AI chip closes the perception-compute-control loop for millisecond-level decisions, with SDK, URDF, and development toolchains. Validated over 300,000+ durability test cycles with automotive-grade drive-and-control boards, the hand delivers the sensing, dexterity, and reliability that long-term embodied AI, robotics research, and high-dynamic tasks demand.
About DexRobot:
DexRobot integrates hardware, perception, data, models, and scenarios into a continuous pipeline: a complete dexterous hand portfolio; native multimodal perception; self-developed data acquisition and models that turn human demonstrations into training data and skills; and real-world scenarios validating every link, closing the loop from hardware to physical tasks, demonstrated by DexGEM on real robots.
For more information, please visit https://www.dex-robot.com/en or contact us at [email protected].
