Aero Hand Open: A Simulation-Ready Tendon-Driven Hand for Dexterous Manipulation Learning

Robotic dexterous manipulation has evolved from simple grippers to complex multi-finger hands, driven by advances in actuation, sensing, and control algorithms. Tendon-driven hands closely mimic human anatomy, offering high dexterity at low cost. However, their complex Cable routing, elasticity, and friction pose significant challenges for accurate simulation and learning. Prior works like Shadow Dexterous Hand and ILDA have achieved high performance but at prohibitive costs or with limited repairability. Recent trends focus on leveraging simulation for policy training, yet the fidelity gap remains a barrier. MuJoCo and other physics engines support rigid body dynamics but struggle with soft, cable-based mechanisms. This paper builds on these efforts, aiming to create a faithful, repairable, low-cost tendon-driven hand with high simulation fidelity, enabling scalable learning.

The main challenge lies in accurately modeling the Cable transmission of tendon-driven hands within simulation environments, capturing complex routing, elasticity, and friction effects. Existing models often simplify or ignore these factors, leading to poor transferability of learned policies. Additionally, designing a low-cost, repairable mechanical hand that can perform diverse grasp types complicates the modeling process. Achieving zero-shot transfer of policies trained solely in simulation to real hardware remains an open problem, especially for underactuated, cable-driven systems with coupled joints and limited proprioception.

The paper introduces several key innovations: 1) a detailed Cable routing model in MuJoCo that incorporates pulley geometries and spatial tendons, 2) a system identification approach deriving linear actuation maps validated across channels, 3) a reinforcement learning pipeline trained entirely in simulation with domain randomization, enabling zero-shot transfer, 4) a low-cost, repairable mechanical design with all parts 3D printed, facilitating maintenance and scalability. These innovations collectively address the fidelity gap and practical deployment challenges, setting new standards for tendon-driven robotic hands.

• �� Mechanical Design: 3D printed, modular, low-cost hand with 16 joints, 7 motors.
• �� Cable Simulation: Modeled each tendon as a spatial tendon wrapping around pulleys, including springs and friction.
• �� System Identification: Measured cable path differences at key poses, derived linear coefficients for each joint.
• �� Simulation Validation: Compared simulated and real cable lengths, joint angles, and forces across multiple tasks.
• �� Policy Training: Used PPO in MuJoCo, with domain randomization on cable routing, spring stiffness, and initial poses.
• �� Deployment: Directly transferred policies to hardware, no fine-tuning needed.
• �� Evaluation: Tested grasp diversity, force exertion, and task speed across real and simulated environments.

Experiments employed MuJoCo simulations with randomized Cable parameters, training policies on 33 grasp types from the GRASP taxonomy. Validation involved real-world tests on the Aero Hand Open, measuring success rate, positional accuracy, and force exertion. The cube rotation task tested manipulation precision. Baseline comparisons included simplified models and traditional control methods. Hyperparameters such as learning rate, domain randomization ranges, and training duration were optimized for robustness. Results consistently showed high transfer success, with quantitative metrics validating the simulation fidelity and policy effectiveness.

The policies trained in simulation achieved over 85% success rate in real-world grasping across all 33 categories, with positional errors less than 3mm. In cube rotation, the system maintained 1.2Hz operation with high accuracy. The actuation map’s error was within 5%, confirming the model’s precision. The mechanical hand demonstrated durability over 400,000 cycles, with cable tensions well within safety margins. These results underscore the effectiveness of the simulation-based training pipeline and the fidelity of the Cable model.

This approach enables low-cost, high-precision robotic hands for industrial automation, service robots, and research platforms. Its repairability and scalability make it suitable for widespread deployment in environments requiring diverse manipulation skills. The zero-shot transfer capability reduces development time, facilitating rapid prototyping and adaptation to new tasks. Future integration with sensory feedback could further enhance performance in complex, unstructured environments.

The linear Cable model may not fully capture nonlinear elasticity and frictional effects, limiting accuracy under extreme conditions. Long-term durability depends on material fatigue, which is not modeled. The reliance on extensive simulation data increases training costs. Future work should incorporate nonlinear models, sensor fusion, and adaptive learning to address these issues, aiming for more robust, versatile systems.