Mechanical Foundations: Underactuation, Compliance, and Object Deformability

Underactuated robotic hands achieve adaptability to object geometry by using fewer actuators than degrees of freedom, relying on passive mechanical elements — springs, tendons, or compliant linkages — to distribute grasping forces across phalanges. This mechanical intelligence reduces control complexity but introduces significant challenges when objects are deformable, because the passive compliance of the hand interacts with the compliance of the object in ways that are difficult to predict. As documented by PatSnap’s innovation intelligence platform, this intersection of underactuated mechanism design, deformable object modelling, and adaptive control theory spans roughly 60 sources across institutions in China, Europe, North America, and Asia-Pacific.

Structural compliance can be introduced either at joints or in finger-pads, and each choice carries different implications for deformable object grasping. Research from the University of Auckland (2019) shows that designs using a single actuator per finger that retain adaptive grasping under significant object pose uncertainty are especially promising for environments where object properties are unknown. These designs conform passively to the object’s geometry — a critical first step before active control can be applied.

Tendon-driven underactuated mechanisms are among the most commonly studied configurations. Work from Stanford University (2022) presents a framework explicitly accounting for friction and dynamic changes in palm width, link lengths, and transmission ratios — parameters that directly affect how much force is safely transmitted to a deformable object without causing damage. This work underscores that geometry variation must be treated as a real-time control variable, not just a design parameter.

Columbia University (2021) demonstrates that using springs as agonists and tendons as antagonists allows the implementation of desired postural synergies — a single motor can generate multiple stable grip configurations, intrinsically distributing contact forces across soft or deformable targets. Meanwhile, research from the Istituto Italiano di Tecnologia (2020) establishes that the resulting stiffness at the grasped object is dominated by the most compliant element in the kinematic chain — typically the underactuated hand itself — and proposes a controller that reshapes the stiffness ellipsoid by combining the hand’s limited contribution with the robotic arm’s larger actuation bandwidth.

Model-Free and Adaptive Control Strategies for Grasp Stability

The fundamental motivation for model-free adaptive control in this domain is that accurate dynamic models of deformable objects are either unavailable at runtime or computationally prohibitive to evaluate in real time. Several complementary strategies have emerged to address this constraint, each suited to different actuator types and sensing configurations.

Fuzzy and Sliding-Mode Hybrid Force Control

Fuzzy force control parameterized by estimated object stiffness represents one of the most accessible model-free approaches. Research from the State Key Laboratory of High Performance Complex Manufacturing, Changsha (2019) presents a fuzzy force controller validated through simulations across six different stiffness levels, confirming that adapting controller behavior based on inferred object compliance leads to effective force regulation without requiring an explicit deformable object model. The fuzzy inference framework naturally handles the uncertainty inherent in stiffness estimation from contact measurements.

A more complete hybrid approach from Universidad Distrital Francisco José de Caldas (2019) uses two distinct sliding-mode control surfaces — one for the slipping state and one for the non-slipping state — unified through a fuzzy inference block. Force sensors on each finger continuously monitor the grasp state, and control transitions between surfaces. Laboratory validation on spherical and cylindrical objects confirms that the minimum force required to prevent slippage is successfully enforced without exceeding safe force thresholds for compliant objects.

Model-Free Adaptive Control with Data-Driven Linearization

A direct application of model-free adaptive control (MFAC) to robotic hands appears in work from Northeastern University, Shenyang (2021), which applies Taylor series expansion and the differential mean value theorem to transform a nonlinear shape memory alloy (SMA)-driven hand system into an equivalent linearization that depends solely on measurement data. Combined with prescribed performance control — which constrains tracking error within a preassigned domain — this method achieves stable force tracking on a hand system whose dynamics are too complex to model explicitly. The SMA actuator’s nonlinearity and hysteresis are representative of the challenges posed by soft or underactuated systems intended to handle deformable objects.

The adaptive model-free control with nonsingular terminal sliding-mode (AMC-NTSM) from Inha University (2020) leverages one-sample delayed measurements to cancel nonlinearities and uncertainties in real time, without relying on a parametric model. The nonsingular terminal sliding variable ensures finite-time convergence of tracking errors — especially valuable in grasp force control where delayed response can cause object damage or slippage.

“Decentralized tactile slip-prediction controllers at individual fingertips produce globally stable grasps without any central communication between finger controllers — local slip prediction at each contact point is sufficient.”

Neural Network and Adaptive Learning-Based Control

Work from Beijing Institute of Tracking and Telecommunications Technology (2021) combines adaptive neural networks with sliding mode control using the Udwadia-Kalaba equation to handle uncertain constrained forces — primarily friction — across all finger links. The neural network estimates friction online, and the combined controller tracks both joint angles and contact forces on a 3-DOF dexterous hand, with chattering suppressed through the adaptive component.

Force-guided grasping of fragile and deformable objects is addressed by Harbin Institute of Technology (2020), where a neural network regression model predicts grip force from surface EMG (sEMG) signals with R² = 0.982. The predicted force command drives an admittance controller that converts the force reference into compliant motion, adapting continuously to the resistance offered by the deformable object. Research from IEEE-published work at Universität Hamburg (2021) extends this further, augmenting a vision-based teleoperation mapping with a biomimetic compliance controller that simultaneously adapts impedance parameters and feedforward force during execution — since neither parameter alone can handle the nonlinear load-deformation relationship of soft objects.

Tactile Sensing and Slip Prevention as Feedback for Adaptive Grasping

Tactile sensing provides the most direct observable signal for detecting imminent slippage and deformation onset during grasping, and its integration with model-free control is a major theme across the surveyed literature. The field has moved from centralized force-control loops toward decentralized, sensor-rich architectures that require no object model whatsoever.

Research from Technische Universität Darmstadt (2020) proposes a modular, biomimetic approach where each fingertip independently runs a slip-prediction controller using BioTac tactile sensors. The paper demonstrates that global grasp stability emerges without any central communication between finger controllers — local slip prediction at each contact point is sufficient. This decentralized architecture is inherently robust to model uncertainty and generalizes to novel objects without retraining.

This approach is extended to hierarchical dexterous manipulation by Max-Planck-Institut für Intelligente Systeme (2020). A higher-level reinforcement learning policy governs manipulation objectives, while low-level tactile-based grip stabilization controllers independently ensure no slip occurs during the execution of each action. The separation between manipulation planning and grip stability enables the reinforcement learning policy to learn task strategies without having to explicitly model contact mechanics — a significant practical advantage when objects deform unpredictably.

For multi-fingered robotic hands, Fuzhou University (2020) presents a combined pipeline: a Deep Neural Network detects contact events and classifies object material from tactile data in real time, while a Gaussian Mixture Model estimates contact force and contact location. These estimates drive a force control loop that stabilizes unknown objects without any prior knowledge of their shape or physical properties — a fully model-free pipeline from sensor to actuation.

Slip detection in prosthetic contexts is addressed by Università Campus Bio-Medico di Roma (2016) through a two-level control architecture: a policy-search learning algorithm at the high level, and a parallel force/position slip-reactive controller at the low level. Tests on the IH2 hand with force sensors demonstrate reliable slip prevention across diverse grasped objects, including deformable ones. Shadow Robotics (2021) extends tactile-driven control to stable 3D pinching, using a deep convolutional neural network operating on optical tactile sensor images to estimate contact surface orientations, then rolling fingertips to reach an equilibrium grasp pose validated on objects of varying softness.

Impedance and Admittance Control for Deformable Object Interaction

Impedance and admittance control frameworks are widely adopted for deformable object grasping because they inherently couple force and position regulation, allowing a robotic hand to remain compliant with respect to the object’s reaction forces while maintaining overall grasp stability. The challenge is that standard fixed-parameter impedance controllers cannot track the continuously evolving nonlinear stiffness of a deforming object — making online parameter adaptation essential.

National Taiwan University of Science and Technology (2020) combines impedance control, fuzzy logic, and iterative learning control into a single framework. The fuzzy impedance controller estimates optimal impedance parameters in real time for unknown objects, while the iterative learning component continuously refines the fuzzy rule base based on accumulated task performance. A real-time slip prevention module adjusts grip force when the object begins to slide during lifting — directly addressing the instability that arises when deformable objects shift under load.

Safe force-controlled grasping of deformable objects such as paper and plastic cups is demonstrated by Aalto University (2020). Resistive force and bend sensors embedded in a pneumatically actuated soft hand feed into a proportional-integral controller that regulates pneumatic pressure to achieve the desired grip force. The data-driven calibration procedure for force estimation eliminates the need for an explicit deformation model, making this approach directly applicable to heterogeneous deformable targets — a practical advantage highlighted by standards bodies such as ISO in their guidance on safe human-robot interaction.

Passive grasp stability — the ability of a grasp to resist disturbances without active control — is rigorously analysed by Columbia University (2020). By incorporating the maximum dissipation principle into the grasp stability model, it becomes possible to distinguish which disturbance directions are passively resisted by non-backdrivable actuators and which require active compensation. This energetic framework is essential for designing the active control layer to target only the disturbances that passive mechanics cannot absorb — particularly relevant for deformable objects whose reaction forces evolve continuously.

A complementary passive analysis framework from Columbia University (2018) formulates the passive reaction problem as a mixed-integer program to account for nonlinear phenomena including joint gear non-backdrivability. This provides a principled methodology for partitioning control authority between mechanical passive compliance and active adaptive controllers — a methodology also referenced in robotics standards work published by IEEE. Variable impedance control for unknown object stiffness is implemented by Myongji University (2021), where a four-finger underactuated hand with Series Elastic Actuators uses a parameter update block to vary impedance in response to measured contact forces, preventing damage to unknown objects with variable stiffness.

Key Players, Innovation Trends, and the Road Ahead

The surveyed literature of roughly 60 sources reveals a clear geography of innovation, with distinct institutional specialisms and converging technical trends. Understanding who is contributing what — and in which direction the field is moving — is essential for R&D strategy in robotic manipulation.

Leading Research Institutions

Columbia University (Department of Mechanical Engineering, New York) contributes multiple foundational works on grasp stability analysis, including passive reaction analysis, energetic constraints for stability, mechanically realisable synergies, and spring-agonist design paradigms. Istituto Italiano di Tecnologia (IIT), Genoa, leads on compliant underactuated hand design and environmental constraint exploitation for grasping. Technische Universität Darmstadt is notable for its biomimetic decentralised tactile feedback approach, while Max-Planck-Institut für Intelligente Systeme and Shadow Robotics push the frontier of hierarchical and tactile-driven dexterous manipulation applicable to soft targets.

Chinese academic institutions — including Northeastern University Shenyang, Harbin Institute of Technology, Fuzhou University, and Nanjing University of Science and Technology — contribute substantially to model-free adaptive control formulations, data-driven force estimation, and tactile-based stabilisation. The volume and technical depth of these contributions reflects the significant investment in robotics research supported by bodies such as WIPO-tracked Chinese patent filings in this space.

Convergent Innovation Trends

Three overarching trends emerge from the literature synthesis. First, there is a clear convergence toward hybrid architectures that combine passive mechanical compliance with active model-free adaptive control — neither alone is sufficient for deformable object grasping. Second, there is increasing adoption of deep learning for contact state estimation from tactile sensors, replacing hand-crafted feature extractors with end-to-end neural pipelines. Third, there is growing use of prescribed performance control frameworks that guarantee bounded tracking error without requiring plant models — a particularly attractive property for systems operating on objects whose mechanical properties are unknown at deployment time.

The PatSnap platform for R&D teams provides landscape analysis tools that can map assignee activity, citation networks, and technology convergence across all four of these control paradigms in real time.