BiDexHand: Open-Source Robotic Dexterous Hand

• BiDexHand is a cable-driven, open-source biomimetic robotic hand platform featuring 16 actuated DoF and five passive couplings that emulate natural human finger movement.
• Its design incorporates tendon-driven actuation with 3D-printed anti-parallelogram linkages and multiple control modalities, including joint-level PID, task-based Cartesian control, and vision-based motion shadowing.
• Validated on all 33 GRASP taxonomy grasps with precise force and motion feedback, BiDexHand supports rapid research replication and complex human-robot interaction experiments.

BiDexHand is a cable-driven, open-source, biomimetic robotic hand platform optimized for research in dexterous manipulation, teleoperation, reinforcement learning, and human-robot interaction. Characterized by 16 independently actuated degrees of freedom (DoF) and five mechanically coupled joints that emulate natural human finger motion, BiDexHand provides validated performance in executing all 33 grasp types of the GRASP taxonomy and enables complex, vision-based, and teleoperated control modalities suitable for benchmark-class experimentation and application (Weng, 20 Apr 2025).

1. Mechanical Architecture and Kinematics

BiDexHand features four identical fingers, each equipped with three actively actuated DoF—distal interphalangeal (DIP), proximal interphalangeal (PIP), and metacarpophalangeal (MCP) flexion/extension—and a fourth mechanically coupled passive joint, realized via a 3D-printed anti-parallelogram (crossed four-bar) linkage. This coupling serves to reproduce the anatomical DIP-PIP coupling seen in humans. The thumb comprises four serially arranged joints: interphalangeal (IP), MCP flexion/extension, MCP abduction/adduction, and carpometacarpal (CMC) flexion/extension, all actively driven. In total, BiDexHand realizes 16 actuated DoF plus five passive couplings, corresponding to 21 joint axes (Weng, 20 Apr 2025).

Cable-driven actuation is achieved via brushless servo motors housed proximally to minimize inertia. Tendon routing is configured in an N-pattern with low-friction PTFE sleeves and antagonistic pull-pull cables. Each actuated joint generates torque $\tau = T r$, where $T$ is cable tension and $r$ is pulley radius. Passive DIP angles $\theta_{D}$ are derived via closed-loop kinematics of the four-bar:

$$\theta_{D} = 2\,\arctan\left(\frac{l_{1}-l_{2}}{l_{1}+l_{2}}\,\cot{\frac{\theta_{P}}{2}}\right),$$

where $\theta_P$ is the actuated PIP angle, and $(l_1, l_2)$ are linkage lengths.

For forward kinematics, each finger is a serial chain modeled via Denavit–Hartenberg (D–H) convention; inverse kinematics resolve commanded fingertip positions while accounting for the mechanically coupled joints. Dynamic modeling employs Lagrangian formalism, incorporating tendon elastic effects and coupling (Weng, 20 Apr 2025).

2. Control System and Feedback Modalities

The control architecture is implemented on an ESP32 microcontroller running a ROS 2 node stack. Three primary control modalities are supported:

• Joint-level Position Control: Each joint is regulated by a PID controller, using high-resolution magnetic encoders ($\approx0.1^\circ$) for feedback.
• Task-based Cartesian Control: High-level motion planners convert task commands (e.g., end-effector velocities) to joint trajectories using inverse kinematics and a resolved-motion-rate controller with Jacobian pseudo-inverses.
• Vision-based Motion Shadowing: Human hand poses are captured via RealSense D405 or VR-based cameras, processed into joint angles $\theta^H$ (human), mapped to robot joint space via a kinematic adapter $f: \theta^H \rightarrow \theta^R$, supporting teleoperation at 30 Hz frame rates.

Telemetry includes 12 magnetic encoders and five fingertip force sensors, enabling real-time monitoring. Closed-loop force/position hybrid control is under development; currently, force feedback is primarily for monitoring (Weng, 20 Apr 2025).

3. Performance Benchmarks and Validation

BiDexHand achieves full coverage of the 33-class GRASP Taxonomy, demonstrating robust performance across power, precision, lateral, spherical, and hook grasps. Success is measured by object retention under 0.5 m/s perturbation and force-distribution analysis (Weng, 20 Apr 2025).

In the Kapandji thumb opposition test (11 poses), BiDexHand attains 9/11 achievable positions, limited by the mechanical CMC linkage’s abduction range. Fingertip normal force peaks at 2.14 N (σ ≈ 0.05 N) per digit, with the hand capable of lifting 10 lb (4.54 kg) objects—implying a distributed tendon tension of 8–10 N per finger.

4. Open-Source Resources and Replicability

Under an MIT license, BiDexHand’s full stack is released at https://github.com/wengmister/BiDexHand. The repository includes:

• CAD/parametric models (SolidWorks, STEP) for all hand components
• ESP32 C++ firmware with ROS 2 communication interfaces
• ROS 2 software packages for high-level control, kinematics/dynamics, and vision pipelines
• Documentation: bill of materials (e.g., Kondo KRS-788HV servos, 0.6 mm Dyneema cables), assembly guides, 3D print settings (polycarbonate, 200 µm layers), and wiring diagrams

Replication is further supported by stepwise assembly instructions and exploded-view imagery (Weng, 20 Apr 2025).

5. Comparative Perspectives and Applications

The BiDexHand architecture represents one approach in the spectrum of dexterous, tendon-driven hands. For example, Dex-Hand 021 employs high-strength tungsten cables, proprioceptive force estimation, and admittance control, achieving 19 DoF and advanced compliance characteristics with a 1 kg total mass (Yuan et al., 5 Nov 2025). The BiDex teleoperation ecosystem integrates BiDexHand (in various versions) into low-latency bimanual systems for human demonstration collection and complex manipulation, emphasizing compatibility with motion capture gloves, rapid inverse kinematics, and ROS (Shaw et al., 2024).

Typical applications include grasping in industrial, service, and research contexts; as a platform for teleoperation; and as a research-grade hand for in-hand manipulation learning and benchmark studies. BiDexHand’s validated performance enables it to serve in GRASP taxonomy tasks, Kapandji benchmarking, vision-based imitation, and as a base for reinforcement learning or learning from demonstration pipelines (Weng, 20 Apr 2025, Shaw et al., 2024, Zhou et al., 2024).

6. Limitations and Future Directions

Identified limitations of BiDexHand include open-loop position control (absent closed-loop force regulation), measurable tendon friction and backlash resulting in ≈5° steady-state error, and a CMC linkage that constrains extreme thumb opposition. Future development directions target closed-loop hybrid control using fingertip sensors, optimized tendon routing and liners for reduced hysteresis, learning-based grasp planning for unstructured environments, and hardware modification to achieve full Kapandji reach (Weng, 20 Apr 2025).

This synthesis strictly adheres to published results and technical disclosures on BiDexHand and close derivatives; all claims, technical details, and metrics are grounded in peer-reviewed arXiv sources.