Gasgoo Munich-TARS took the wraps off its TARS DexHand dexterous hand on June 1 in Austria, marking the product's global debut at the International Conference on Robotics and Automation (ICRA), according to Gasgoo.In terms of hardware, the DexHand employs a 21-degree-of-freedom quasi-direct drive architecture, replicating the skeletal structure and joint distribution of a human hand at a 1:1 ratio. This design enables human-level capabilities in joint coordination and complex, high-precision movements—while maintaining smoothness, accuracy, and consistency in manufacturing.Image Credit: TARSFrom a data perspective, the DexHand aligns deeply with a "Human-centric" data paradigm. It achieves high-fidelity mapping of human motion data onto the robotic hand, significantly boosting the utilization of embodied intelligence data.On the modeling front, the DexHand integrates with Tashi’s general embodied large model, AWE 3.0. Leveraging the newly released TacForeSight technology, the system channels wrist force and torque signals into OmniVTA, a visuo-tactile world model.This allows the DexHand to perceive physical textures—rough, soft, or hard—and even detect microscopic details as fine as 0.03mm. Beyond sensing, it actively predicts changes in the physical environment and adjusts its strategy accordingly. By compressing perception, understanding, prediction, and manipulation into a single closed loop, the system achieves true "hand-brain integration."Founded in February 2025, TARS is led by founder and CEO Chen Yilun, formerly Huawei’s CTO for autonomous driving and chief scientist at its Car BU. Chairman Li Zhenyu, meanwhile, previously served as president of Baidu’s Intelligent Driving Group (IDG), where he spearheaded the Apollo open platform and the Apollo Go robotaxi service.On the hardware side, Tashi has already rolled out its A-series wheeled industrial robots and T-series bipedal general-purpose robots. The company has secured a string of technical breakthroughs in Human-centric data collection, general embodied intelligence models, and precision assembly of flexible wiring harnesses. Recent milestones include the release of the AWE 3.0 large model and the SenseHub data acquisition kit, as well as the open-sourcing of the "WIYH" embodied VLTA multimodal dataset.Beyond showcasing the DexHand, the company demonstrated its A1 robot performing fine-grained, daily-life packing tasks powered by AWE 3.0. Throughout the day, the robot repeatedly executed complex, multi-step sequences—grabbing pencils, placing them in cases, zipping bags, and organizing backpacks—completing continuous long-horizon tasks involving seven or more steps.
+5 others
