Clutch: Mechanisms in Robotics & Engineering

• Clutch is a controllable coupling that regulates motion, force, and torque across diverse engineering systems, including automotive powertrains and robotic applications.
• In automotive and drivetrain contexts, clutches switch power delivery modes by transitioning from slip to locked states, enhancing energy efficiency and control.
• Electrostatic and variable-impedance clutches enable precise torque modulation and low-power actuation in robotics and soft actuator systems.

A clutch is a controllable coupling, locking, or resistive interface that regulates relative motion, force, or torque between subsystems. In the contemporary research literature, the term spans rotary and linear drivetrain elements, passive locking mechanisms for springs and transmissions, electroadhesive and electrostatic friction interfaces for robotics, magnetorheological and pneumatic variable-impedance devices, and clutch-mediated control architectures in soft and elastic robots. The same term also appears as a title or acronym in adjacent fields, where it denotes formal systems or computing methods rather than mechanical hardware (Amish et al., 2023, Kim et al., 2022, Kong et al., 18 Jun 2025, Gong et al., 2023, Gregersen et al., 2023, Tokuda et al., 22 Jun 2026).

1. Fundamental mechanical role

In the surveyed engineering literature, a clutch is defined less by a single geometry than by its function: it either couples and decouples a power path, holds a load without continuous actuation, or modulates impedance by varying interfacial resistance. This includes binary engagement in automotive drivelines, continuously adjustable torque transfer in electroadhesive rotary devices, and one-direction self-locking in capstan mechanisms. A recurring design distinction is between holding and driving: some clutches primarily preserve state with near-zero holding power, whereas others are used as tunable transmission elements (Feizi et al., 2022, Gong et al., 2023, Morselli et al., 10 Apr 2026).

The capstan family makes this functional viewpoint explicit. When a flexible element wraps around a drum, the holding-tension ratio obeys the classical exponential law

$T_{\mathrm{out}} = T_{\mathrm{in}} e^{\mu \theta},$

where $T_{\mathrm{in}}$ and $T_{\mathrm{out}}$ are the low- and high-side tensions, $\mu$ is the friction coefficient, and $\theta$ is the wrap angle. For rotary use, the corresponding torque is typically

$$\tau = R\left(T_{\mathrm{out}} - T_{\mathrm{in}}\right).$$

This relation explains why wrapped interfaces can transform modest local traction into large global holding force, and why clutch performance can depend more strongly on geometry than on area alone (Kim et al., 2022).

Automotive and powershift studies formalize the same idea in slip coordinates. In a two-speed powershift, clutch synchronization is represented by relative speeds such as

$$\delta_a(k)=w_1(k)-R_a w_2(k), \qquad \delta_b(k)=w_1(k)-R_b w_2(k),$$

and engagement corresponds to driving these slips to zero under torque limits. This formulation treats the clutch as the hybrid element that changes system mode from slip to synchronized lock, or, when both paths are constrained simultaneously, to a full-lock state (Morselli et al., 10 Apr 2026).

2. Electrostatic and electroadhesive clutches

Electrostatic and electroadhesive clutches use electrically generated normal force to create controllable frictional shear. In the literature surveyed here, they appear in planar sliding-film form, in rotary disc form, and in capstan-wrapped form. Their chief attractions are low mass, low holding power, and compatibility with thin, compliant robotic structures, but their performance depends strongly on dielectric behavior, surface condition, and the coupling between normal pressure and interfacial friction (Amish et al., 2023, Feizi et al., 2022).

The highest reported specific shear stress in the provided material appears in the Johnsen–Rahbek capstan clutch. By combining the Johnsen–Rahbek electroadhesion effect with capstan amplification, that system generated 31.3 N/cm$^2$ shear stress and 7.1 Nm total holding torque while consuming only 2.5 mW/cm$^2$ at 500 V; it also reported that large-angle designs with $\theta > 2\pi$ are more efficient than planar or small-angle designs with $T_{\mathrm{in}}$0, and identified polybenzimidazole (PBI) as the first unfilled polymeric material reported to exhibit the Johnsen–Rahbek effect (Amish et al., 2023).

Planar electrostatic clutches remain important where flexibility and low thickness dominate. In an antagonistic musculoskeletal joint using HASELs, a planar electrostatic clutch built from 125 $T_{\mathrm{in}}$1m PET, 50 nm aluminum electrodes, and a 6 $T_{\mathrm{in}}$2m PVDF terpolymer dielectric achieved 4.25 kg holding force at 100 V and 8.41 kg at 150 V, corresponding to 2.8 N/cm$T_{\mathrm{in}}$3 and 5.6 N/cm$T_{\mathrm{in}}$4 shear stress, with 5 ms locking time and 15 ms release time. In that architecture, the clutch functions as a series tensile link when engaged and an extensible branch when disengaged, preventing displacement loss from tendon slack and enabling smooth antagonistic motion up to 3.2 Hz (Kazemipour et al., 2024).

Rotary electroadhesive disc clutches pursue a different objective: continuously adjustable transmitted torque. A smart torque-adjustable rotary electroadhesive clutch for human–robot interaction modeled annular torque transfer by

$T_{\mathrm{in}}$5

with shear stress derived from electroadhesive normal pressure and friction. The paper compared three plate-pair configurations and found that a dielectric-against-steel interface driven by AC produced the smoothest output, around $T_{\mathrm{in}}$6 N·m, while also addressing DC polarization-induced degradation through alternating-waveform activation. One disc pair weighing about 20 g transferred up to 3.9 N·m, and the system reported a torque-to-power ratio six times better than commercial magnetic particle clutches (Feizi et al., 2022).

In soft robotics, electroadhesive clutches are often used not as shaft couplers but as active strain limiters. Patterned electroadhesive clutches mounted on a single-chamber soft pneumatic actuator produced pyramidal, round, and plateau shapes, enabled five directional manipulation modes, and applied forces up to 3.2 N during rapid clutch deactivation (Campbell et al., 2022). A later elastomeric implementation modeled clutch force as

$T_{\mathrm{in}}$7

and reported a silicone-sheathed electrostatic clutch with about 25 N maximum tensile holding force at 10 mm extension, including approximately 22 N from the clutch and 3 N from the sheath (Campbell, 3 Apr 2026).

3. Capstan clutches, passive locking, and mechanical multiplexing

Capstan clutches exploit exponential friction amplification along a wrapped interface. Their central advantage is that a very small control or preload force can hold a much larger load once the wrap self-tightens, which makes them attractive for passive locking, energy accumulation, and compact force routing in robotic systems (Kim et al., 2022).

A clear example is the lockable compression spring. There, a small capstan clutch passively locks a mechanical spring, enabling energy storage at arbitrary deflection. The device could lock over 1000 N force, unlock in less than 10 ms, and do so with a control force less than 1% of maximal spring force. Its clutch-specific design equation was reported as

$T_{\mathrm{in}}$8

with $T_{\mathrm{in}}$9 mm, $T_{\mathrm{out}}$0 mm, $T_{\mathrm{out}}$1 mm, $T_{\mathrm{out}}$2 mm, and $T_{\mathrm{out}}$3. The implemented preload was 0.65 N, unlock energy was less than 0.1 J, and measured round-trip energy efficiency was 74–84%, with an average around 80% (Kim et al., 2022).

Electrostatic capstan clutches extend the same wrapped-interface logic to electrically controlled multiplexing. In a mechanical multiplexer, a rotary Johnsen–Rahbek electrostatic capstan clutch used a 25.4 mm hollow stainless-steel shaft coated with 55 $T_{\mathrm{out}}$4m PBI and a 10 mm stainless-steel band wrapped through 3.54 radians. The clutch transmitted 0.43 N·m at up to 60 rpm, corresponding to 2.70 W in the reported lifting test, and an average engagement time of 481 ms. Because each capstan clutch is one-directional, the multiplexer used two clutches per output—one on a clockwise shaft and one on a counterclockwise shaft—to achieve bidirectional actuation and passive hold via self-locking leadscrews. The system actuated a 4-DoF robotic hand with a single motor and could drive 22.24 N per output up to 5 cm (Amish et al., 14 Jan 2025).

A recurring implication of capstan-based work is that “clutch” can denote a highly architecture-dependent amplifier rather than a simple interface whose capacity scales linearly with area. The literature therefore treats wrap angle, band compliance, pretension, and one-way behavior as first-order design variables, not secondary packaging details (Amish et al., 2023).

4. Semi-active and variable-impedance clutches

A second major research direction uses clutches as semi-active impedance elements. In these systems, the clutch does not primarily transmit power continuously; rather, it modulates resistance, damping, or locking torque to shape interaction forces. Magnetorheological and pneumatic clutches dominate this category (Kong et al., 18 Jun 2025, Jiang et al., 2021).

In an upper-limb teleoperation exoskeleton, a magnetorheological clutch built with a dual-bearing “MR bearing – coil – MR bearing” structure served as the core haptic actuator. The clutch reached 42.12 N·m at 1.3 A, with 0.2 N·m idle torque, and reported a torque-to-mass ratio of 93.6 N·m/kg, a torque-to-volume ratio of 4.05 $T_{\mathrm{out}}$5 N·m/m$T_{\mathrm{out}}$6, and a torque-to-power ratio of 4.15 N·m/W. Because residual magnetization delayed release, the system used an alternating-field demagnetization strategy. Here the clutch is explicitly semi-active: it resists operator motion but does not inject mechanical energy (Kong et al., 18 Jun 2025).

A hand-exoskeleton variant pursued linear holding force instead of rotary torque. Its miniature linear magnetorheological grease clutch used a micro roller enhancing structure, or MR Contact Pair, in which one roller interacting with hardened magnetorheological grease generated resistance through squeeze-strengthened force $T_{\mathrm{out}}$7, shear-related force $T_{\mathrm{out}}$8, and friction $T_{\mathrm{out}}$9. Each clutch contained 80 MRCPs in two groups. Reported peak holding force was approximately 380 N, and the paper gave a headline force-to-power ratio of 256.75 N/W, while also reporting 276.18 N/W at 2.0 V and 127.05 N/W at 3.0 V. The same paper reported a total exoskeleton support force of approximately 419.79 N for gripping and a total device mass of 1.12 kg, with 0.61 kg attributable to the four clutches (Li et al., 20 Mar 2025). The numerical reporting is not fully internally consistent across sections, but the general conclusion is that the design prioritizes very high holding force at low electrical power.

Pneumatic clutches address a similar problem by pressure-controlled friction rather than field-responsive rheology. A compact pneumatic clutch with integrated sensing used a positive-pressure air pouch to press a TPU-fabric strip against a high-friction layer, achieving about a 24-fold change in impedance force and a maximum force density of 15.64 N/cm$\mu$0. The same device embedded a position-sensing trace with 1 mm resolution and improved soft-gripper gripping force by 73.3% (Jiang et al., 2021). This suggests a broad category of “clutch” in which the primary output is controllable resistance rather than active drive.

5. Automotive and drivetrain clutches

In automotive systems, clutches remain canonical hybrid elements because they switch powertrain topology. The reviewed literature treats clutch state as a discrete variable that changes the governing equations of torque flow, engine speed, and synchronization, which makes clutch control inseparable from overall energy management and shift logic (Gong et al., 2023, Morselli et al., 10 Apr 2026).

For a series-parallel plug-in hybrid electric vehicle, the clutch state $\mu$1 determines whether engine torque is routed mechanically to the wheels or electrically through the generator. The control-oriented model writes

$\mu$2

$\mu$3

$\mu$4

These equations show that clutch engagement locks engine speed to vehicle speed and enables direct drive, whereas disengagement frees the engine to operate on its optimal economic line and routes power to the generator. A continuous-discrete reinforcement-learning strategy that jointly optimized engine torque and clutch state improved energy efficiency by 8.3% over charge-depleting/charge-sustaining control at high SOC and by 4.1% at low SOC, while staying 6.6% and 3.9% above dynamic programming, respectively (Gong et al., 2023).

Two-speed transmission studies emphasize clutch engagement dynamics rather than energy management. A centrifugal-clutch transmission analysis examined a commercial three-shoe centrifugal clutch in two gearbox configurations, reporting that Configuration A produced smoother upshifting while Configuration B gave more responsive downshifting. The same study trained a deep neural network to predict clutch engagement from spring preload and shoe mass, using simulation-generated labels rather than experimental data (Lin et al., 2024).

For powershift control, the clutch appears as a discrete-time synchronization constraint. A two-speed powershift model computed the exact clutch torque required to achieve single-clutch engagement or simultaneous full lock in one sampling step, using the slip variables $\mu$5 and $\mu$6 defined above. The method was designed specifically for real-time simulation and embedded control, replacing continuous-time event detection with sample-wise algebraic torque calculations plus saturation checks (Morselli et al., 10 Apr 2026). A common misconception is that drivetrain clutches are adequately represented by static torque limits alone; this body of work instead treats them as hybrid mode-switching devices whose timing and locking logic are part of the system dynamics.

6. Robotic sequencing, interface signals, and terminological extensions

Robotics research also uses clutches as internal scheduling devices that regulate when elastic energy is stored or released. In a clutched-elastic robot, a contact-implicit optimal-control formulation optimized both continuous control input and clutch sequence simultaneously, avoiding pre-specified mode schedules. On a double pendulum with two Bi-Stiffness Actuators, the optimized clutch schedule achieved a final end-effector speed of 1.5 m/s, compared with 1.1 m/s for a previously guessed sequence, while a switching penalty discouraged unnecessary clutch transitions (Ossadnik et al., 2024). This suggests that, in elastic robots, clutch timing is itself a control variable rather than merely a hardware implementation detail.

At the human–machine interface level, “clutch” can denote a console-mode input rather than an actuator. The Comprehensive Robotic Cholecystectomy Dataset records the da Vinci clutch pedal as a Boolean state and uses kinematic windows to predict clutch usage. In that dataset, the clutch class distribution was highly imbalanced—205,406 not-pressed samples versus 934 pressed samples before undersampling—and the best reported window-level F1 came from LightGBM at 0.8427, with Random Forest close behind at 0.8379 (Oh et al., 2023). Here the clutch is a teleoperation state-transition signal, not a power-transmission mechanism.

The term also extends beyond mechanics altogether. “Clutch” names a higher-order probabilistic relational separation logic that introduces asynchronous probabilistic couplings via presampling tapes and proves contextual refinement and equivalence for rich higher-order probabilistic programs (Gregersen et al., 2023). In processing-using-DRAM, “Clutch” expands to Comparison Algorithm using Lookup Table with Chunked Temporal Coding, a vector–scalar comparison method that improved end-to-end application throughput and energy efficiency by an average of 12× and 69× over optimized CPU/GPU execution, and by 2.9× and 3.0× over bit-serial processing-using-DRAM (Tokuda et al., 22 Jun 2026). In hand-motion generation, CLUTCH denotes a language-model-based system for text-conditioned in-the-wild 3D hand motion, supported by the 3D-HIW dataset of 32K sequences and a tokenizer called SHIFT (Thambiraja et al., 19 Feb 2026). Accordingly, a strict equation of “clutch” with mechanical shaft coupling no longer matches current research usage; the term now denotes a broader class of controlled gating, coupling, or comparison mechanisms, depending on disciplinary context.