DLR-Hand II: Tactile Dexterity in Robotics

• DLR-Hand II is a multi-fingered robotic hand with 12 active joints featuring tendon-driven design for advanced tactile in-hand manipulation without reliance on external sensors.
• It employs reinforcement learning with sim2real transfer, using proprioceptive feedback and domain randomization to achieve robust, continuous manipulation tasks.
• The system integrates self-contained calibration and sensorized soft skins to enhance tactile feedback, ensuring precise grasping and adaptive control under varying conditions.

The DLR-Hand II is a multi-fingered, torque-controlled humanoid robotic hand designed for advanced dexterous manipulation, robotics research, and practical deployment in challenging settings that demand robust, adaptive and highly controlled in-hand manipulation. It has emerged as a canonical testbed for tactile in-hand control, sim2real transfer of learning-based policies, self-contained calibration, and continuous manipulation tasks without reliance on external sensing modalities. The following sections detail its physical characteristics, sensor suite, control architectures, leading results, learning paradigms, calibration schemes, and future implications.

1. Mechanical Design and Sensing Capabilities

DLR-Hand II comprises four fingers, each with three actuated degrees of freedom and one yielding passive joint, resulting in 12 active joints in total. The fingers are tendon-driven with rolling-contact joint mechanics, supporting complex multi-contact grasps and force closure. Integrated torque sensors on the output side of each joint enable sensitive detection of contact and precise force control. Joint measurement offsets, drivetrain elasticity (parameterized by $K_e$), and slip-stick friction are explicitly modeled in both software and hardware deployments.

Tactile feedback is obtained exclusively from internal joint position and torque sensors. This configuration eliminates dependency on vision or external markers and facilitates closed-loop control in the absence of global state information.

2. Reinforcement Learning for Purely Tactile Manipulation

A defining achievement on the DLR-Hand II is the execution of purely tactile, continuous in-hand manipulation tasks such as robust cube rotation, reorientation in $\frac{\pi}{2}$ raster patterns, and speed-conditional object manipulation (Sievers et al., 2022, Pitz et al., 2023, Pitz et al., 20 Nov 2024, Röstel et al., 2023). Policies are trained via deep RL in simulation (PyBullet or IsaacSim) and transferred zero-shot to hardware, leveraging accurate system identification and domain randomization (e.g. joint angle offsets, friction variation, sensor noise, “sticky actions”). Control inputs are filtered at high frequency (1 kHz in hardware, 500 Hz in sim) using low-pass filters with specified cutoff (e.g. 2 Hz), protecting hardware from unstable commands.

Architecturally, controllers use network policies (generally fully connected with two layers of 512 units) conditioned on proprioceptive windows (e.g., 0.5 seconds of past joint/torque readings) and goal orientation. Advanced modular schemes decouple control and state estimation: state estimators (differentiable particle filters or learned networks) reconstruct object pose and orientation from proprioceptive data, feeding state estimates into goal-conditioned policies.

Estimator-coupled RL further improves robustness by training the estimator and controller together, so the policy “learns” to avoid ambiguous tactile scenarios and acts conservatively under state uncertainty (Röstel et al., 2023). Performance metrics include over 46 consecutive full cube rotations, success rates up to 99% in $\frac{\pi}{2}$ raster reorientation, and the fastest speed-adjustable, vision-free tactile manipulations reported (Pitz et al., 20 Nov 2024), with robust sim2real transfer and resilience to disturbances (e.g. variation in cube size, finger pulling).

3. Grasping: Multi-Contact, Power Grasps, and Learning-based Controllers

DLR-Hand II supports arbitrary grasp types—including fingertip, multi-contact, and full power grasps—without explicit contact mapping (Winkelbauer et al., 18 Sep 2024). Grasp controllers estimate external wrench applied to the object using measured joint torques and predict the corrective joint commands needed to compensate, via an impedance control loop. The contacts are modeled elastically, and the controller solves for minimal torque corrections that stabilize the object’s pose under external disturbance.

Two neural networks (wrench estimator and torque predictor) are trained in a supervised manner on data synthesized from analytic grasp models, enabling real-time operation (6 ms control cycle) and robust adaptation to unknown objects using only coarse 3D shape models (48³ voxelgrids). In simulation, 83.1% grasp stability is maintained under external wrenches up to 10 N, outperforming baselines both in torque efficiency and minimization of involuntary object movement.

4. Calibration: Pairwise Contact-based, Self-Contained Procedure

Kinematic calibration of the DLR-Hand II leverages only internal sensing and pairwise finger contacts (Tenhumberg et al., 2023). The process iteratively moves active fingers toward passive ones until contact is detected by thresholding torque signals (e.g., $\tau_t = 0.1$ Nm). The known zero distance between modeled fingertip geometries at contact allows formulation of a least-squares optimization problem for correcting Denavit–Hartenberg parameters. Global calibration is assured through optimal experimental design (task D-optimality) for selecting informative joint trajectories.

Sensitivity analysis shows that all relevant quantities for dexterous manipulation (relative fingertip difference vectors) are identified with accuracy equivalent to visual tracking-based calibration. Experiments demonstrate reduction of maximal fingertip position error from 17.7 mm uncalibrated to 3.7 mm calibrated (mean error < 1 mm), with only 9 minutes required to collect sufficient contact data.

5. Soft Sensorized Skin Integration and Tactile Enhancement

Recent developments offer bioinspired, sensorized soft skins for dexterous hands, with a 1 mm–thick silicone origami design conforming to the hand’s geometry (Egli et al., 30 Apr 2024). Integration into the DLR-Hand II involves CAD adaptation, multi-material 3D printing or casting, and embedding distributed piezoresistive sensors (e.g., 46 per hand) on flexible PCBs beneath the skin. The skin enhances contact area, grip firmness, pressure mapping, and safety by emulating human skin compliance. Calibration of nonlinear sensor responses, improvement in adhesion, and integration with joint-level controllers are ongoing challenges and research directions. Compared to traditional DLR-Hand II tactile sensors (which are sparse and hard-mounted), the soft skin design offers more uniform, distributed feedback for adaptive grasp and collision-sensitive manipulation.

6. Benchmarking and Data Collection for Dynamic Dexterous Grasping

Benchmark datasets such as DexH2R (Wang et al., 29 Jun 2025) are critical for evaluating dynamic human-to-robot handovers and teleoperated naturalistic grasping on dexterous robot hands in real-world scenarios. The datasets contain thousands of annotated grasp trials, diverse object geometries, and multi-modal sensor streams (multi-view RGB, depth, egocentric wrist views, high-res point clouds), facilitating training and comparison of generative grasp models (cVAE, auto-regressive, and diffusion-policy techniques) for grasp pose preparation, approach trajectory generation, and alignment.

Applying such approaches to DLR-Hand II could further improve adaptive trajectory planning and collision-safe execution. Evaluation reveals that diffusion-policy methods using 3D point clouds offer an effective balance between success rate and safety (collision avoidance) compared to more aggressive auto-regressive methods.

7. Future Directions and Research Implications

DLR-Hand II continues to support research into estimator-coupled RL, multi-contact grasping, tactile-based closed-loop control, self-contained calibration, and adaptive sensorized skins. Promising directions include:

• Extension to objects with arbitrary, unknown geometries
• Incorporation of distributed tactile sensor arrays and soft skins
• Adaptive sim2real domain randomization for contact modeling
• Integration with rich benchmark datasets (e.g. DexH2R) for dynamic grasping
• Uncertainty-aware and hybrid model-based/model-free architectures
• Robust RL for industrial and collaborative settings with high task variability

These advances position DLR-Hand II as an archetype for studies in tactile dexterity, robust in-hand manipulation, and real-world deployment of learning-based robot control strategies.