BRL/Pisa/IIT SoftHand: Adaptive Robotic Grasping

• BRL/Pisa/IIT SoftHand is an anthropomorphic, underactuated, tendon-driven robotic hand that employs human-inspired postural synergies for adaptive and robust grasping.
• It combines compliant mechanical design with 3D-printed modular phalanges and biomimetic tactile sensors to achieve safe and efficient manipulation in dynamic environments.
• Its evolution into educational and control-augmented variants demonstrates cost-effective replication and advanced closed-loop tactile feedback techniques for improved grasp performance.

The BRL/Pisa/IIT SoftHand refers to a family of anthropomorphic, underactuated, tendon-driven robotic hands originating from the Pisa/Istituto Italiano di Tecnologia (IIT) Adaptive Synergy Hand, expanded by the Bristol Robotics Laboratory (BRL) and collaborators into a suite of 3D-printed, tactile-sensor-enabled, educational, and control-augmented variants. This lineage is characterized by the use of human-inspired postural synergies, extreme mechanical underactuation (typically one or two motors for all joints), soft mechanical compliance, and integrated biomimetic optical tactile sensing for robust, adaptive grasping in real-world manipulation tasks. In the following sections, essential aspects of the system design, synergy principles, control architectures, tactile sensing approaches, quantitative performance, and benchmarks are synthesized based on recent literature.

1. Anthropomorphic Mechanical Structure and Synergy Modeling

The canonical Pisa/IIT SoftHand is a five-fingered, anthropomorphic hand comprising 19 joint axes (MCP, PIP, DIP for each finger, plus ab/adduction for the thumb) actuated by a single electric motor via a continuous tendon routed throughout all fingers. Finger joints consist of rigid links with passive compliance implemented by torsional springs or flexural elements, connected using dislocatable, overload-protected pin joints to provide mechanical safety. As the motor reels in the tendon, fingers close simultaneously, and contact with objects halts further closure of individual joints, enabling the remaining joints and digits to continue enveloping the object. This simultaneous, adaptive closure implements the first principal component of human hand posture covariance—the dominant hand "synergy" observed in power grasp studies—quantitatively represented as $q \approx \bar q + P_1 s_1$, where $P_1$ is the first synergy vector and $s_1$ is a scalar synergy amplitude controlled directly by motor rotation (Cairnes et al., 2023).

Subsequent SoftHand designs including BRL/Pisa/IIT derivatives, the SoftHand-A, and 3D-printed variants employ similar principles. Some extend actuation to two motors and antagonistic tendon pairs to enable partial control of individual finger joints (particularly distal or proximal interphalangeal joints), effectively implementing additional postural synergies and permitting selective grasp pre-shaping or finger spreading (Li et al., 2024, Lepora et al., 17 Oct 2025).

2. Underactuation, Compliance, and Tendon Routing

All SoftHand models share extreme underactuation, with one actuator typically controlling 19 DOFs (or two actuators for enhanced selectivity in newer models). Tendon routing is central: a continuous cable or tendon is passed through all phalanges, interacting with pulley geometries and local joint compliance to define the closure sequence. The key kinematic mapping is

$\Delta\ell = \sum_i r_i \Delta\theta_i,$

where $r_i$ are tendon moment arms and $\Delta\theta_i$ joint angles. Joint stiffnesses $k_i$ are tuned so proximal joints lock earlier than distal ones, yielding an adaptive wrapping mechanism. Dislocatable joints and passive compliance ensure that overloads yield mechanically rather than transferring excessive forces to the actuator—for both safety and shape-conformant grasping (Cairnes et al., 2023, Li et al., 2022).

In low-cost, open-source variants, such as the BRL/Pisa/IIT SoftHand, kinematic chains are realized through modular 3D-printed phalanges with gear-pair linkages, polymer bearings for controlled tendon friction, and series elastic elements (rubber bands, springs) introducing additional compliance ("soft synergy") (Li et al., 2022). Pretensioning ensures backlash-free closure, and geometric constraints prevent over-folding.

3. Control Architectures: Synergy, Tactile Feedback, and Adaptive Grasping

The original SoftHand control is minimalist: motor position commands set the synergy amplitude, optionally augmented by motor current sensing for crude force estimation. There is no multi-finger closed-loop control; all adaptability is mechanical (Cairnes et al., 2023).

Recent generations exploit optical biomimetic tactile sensors for closed-loop grasping. The BRL TacTip sensors, as well as miniaturized microTac variants, integrate optical cameras within compliant fingertip domes that image arrays of internal pins ("dermal papillae"), transducing contact normal and shear forces and precise geometry at finger-object interfaces (Lepora et al., 2021, Ford et al., 21 Mar 2025). Tactile images are processed to compute deformation metrics (e.g., the Structural Similarity Index Measure, SSIM), which serve as feedback signals for incremental motor control:

$$e_\text{SSIM}(I) = 1 - \text{SSIM}(I, I_{\text{ref}})$$

with proportional or PI control updating motor setpoints to stabilize contact force and depth.

Deep learning pipelines further enable estimation of contact pose and force for each finger from tactile images. Parallel inference on distributed compute modules (e.g., Jetson Nano, Raspberry Pi) yields full-hand tactile feedback at rates up to 286 Hz. Shear-based control frameworks regulate grasp force dynamically by monitoring shear-error signals $\Delta F_x$, $\Delta F_y$ across consecutive frames, modulating grasp motor commands via PID loops for stable manipulation even under dynamic object forces (Ford et al., 21 Mar 2025, Ford et al., 2023).

Multi-finger tactile feedback supports advanced manipulation modes including slip detection, adaptive reflexes, and human-in-the-loop leader-follower control.

4. Tactile Sensing: Integration, Signal Processing, and Contact Modeling

SoftHand tactile integration centers on the TacTip and microTac sensor families. Each sensor comprises a compliant, hemispherical skin filled with transparent gel over an array of internal elastomeric pins, with a miniature camera tracking pin markers' displacements caused by contact. Marker movements encode normal and shear deformations, supporting pose (x, z, roll, pitch, yaw) and force estimation. SSIM-based deformation or density-based marker position clustering (Gaussian kernel density) are used to detect contact and slippage (Li et al., 2024, Lepora et al., 2021).

Contact force is often estimated by simple spring models,

$P_1$0

where $P_1$1 is marker displacement and $P_1$2 is a calibrated effective stiffness. Deep neural networks are trained with ground truth force and pose data to achieve sub-millimeter and sub-Newton accuracy, enabling robust, high-resolution tactile feedback in real time (Ford et al., 21 Mar 2025).

5. Quantitative Performance and Experimental Validation

Benchmarking of SoftHand variants is extensive in recent work. The BRL/Pisa/IIT SoftHand achieves maximal grasp forces of 19.8 N (total, for objects up to 1 kg) and 5.5 N at a single finger, with typical multi-finger (4-5 fingers) contact on medium/large items. Grasp success rates reach ~80% for medium/large objects, but remain <40% for thin tools due to fingertip geometry limitations. Durability is demonstrated by >200 open/close cycles without component wear (Li et al., 2022).

Tactile-enabled systems (microTac, TacTip) reach SSIM-based closed-loop grasp stability within 1-3 seconds; PID-regulated force feedback adjusts grip reactively to dynamic disturbances while protecting delicate objects from over-gripping (e.g., maintaining normal force below 9 N on paper cup tests). Tactile sensing loops operate at 60–286 Hz, far exceeding previous multi-fingered tactile systems (Ford et al., 21 Mar 2025, Ford et al., 2023, Lepora et al., 2021).

Antagonistic tendon designs (SoftHand-A) support independent control of DIP/PIP joints, adaptive slip correction, and human-hand mirroring, with peak fingertip forces >10 N and contact localization to ~1 mm (Li et al., 2024, Lepora et al., 17 Oct 2025).

6. Comparative Benchmarks and Application Contexts

Relative to fully actuated hands (Shadow Dexterous), SoftHand architectures offer radical mechanical and control simplicity—trading precise finger-by-finger manipulation for enveloping, adaptive grasps. Compared to SDM and iHY hands, SoftHand designs leverage rigid linkages with compliant joints for improved structural strength and higher DOF underactuation (1 motor for 19 DOFs versus 5/8 in iHY). Anthropomorphic geometry confers advantages in tasks requiring lifting, reorienting, and passing items over universal jamming/vacuum grippers, which lack dexterous manipulation capability (Cairnes et al., 2023).

Educational SoftHand-A variants demonstrate that hands-on construction and synergy control concepts can be faithfully reproduced using standard LEGO Mindstorms parts, maintaining adaptive grasping and basic force/kinematic performance (~1.8 N fingertip closing force) suited for teaching (Lepora et al., 17 Oct 2025). Open-source 3D-printed models enable rapid and affordable replication and customization (<£70 parts cost).

Application domains span autonomous grasping, prosthetic devices (leveraging tactile feedback for afferent sensing), gentle human–robot handover collaboration, and leader–follower manipulation tasks (Ford et al., 2023, Lepora et al., 2021).

7. Design Trade-Offs, Limitations, and Future Directions

Key strengths of the SoftHand lineage include anthropomorphic form factor, universal shape adaptation via synergy, intrinsic safety by compliance and mechanical dislocatability, low actuation cost, and simplicity of integration. Limitations persist: grasp modalities are confined primarily to enveloping power grasps and lack native precision posture control or force/torque sensing per finger. Performance on thin or small objects is constrained by geometry, while maximal fingertip forces are bounded by underactuation (Cairnes et al., 2023, Li et al., 2022).

Recent advances—antagonistic tendon mechanisms, closed-loop tactile control, deep learning pose/force estimation—extend the functional envelope toward selective joint actuation, slip-resistant manipulation, and interactive grasping. The open-source release of tactile-enabled, 3D-printable, antagonistic designs, sensor integration pipelines, and educational platforms further democratize research and deployment (Li et al., 2024). Prospective developments focus on multi-synergy architectures, real-time high-bandwidth tactile feedback, and robust manipulation in cluttered or dynamic environments.

---

Table: Core Features of the BRL/Pisa/IIT SoftHand and Major Variants

Variant	Actuation Architecture	Tactile Sensing Integration
Pisa/IIT SoftHand	1 motor, 19 DOFs, single tendon	None (motor encoder/motor current)
BRL/Pisa/IIT SoftHand	1 motor, modular 3D-printed, multi-tendon	Optional TacTip/microTac fingertips
SoftHand-A (Antagonist)	2 motors, antagonistic tendons, joint selectivity	Fully 3D-printed tactile fingertips
Educational SoftHand-A	2 motors, LEGO clutch-differential	None (educational constructs only)

All claims, performance metrics, and technical descriptions are traceable to the cited papers (Cairnes et al., 2023, Li et al., 2022, Lepora et al., 2021, Lepora et al., 17 Oct 2025, Li et al., 2024, Ford et al., 2023, Ford et al., 21 Mar 2025).