Soft Everting Robots: Design and Applications

• Soft everting robots are compliant systems that extend by inverting pressurized body material, allowing low-friction navigation in cluttered spaces.
• They leverage diverse architectures—such as tip-extending vine robots and toroidal designs—for controlled growth, precise steering, and reliable retraction.
• Advanced techniques like variable stiffness, finite element modeling, and innovative tip interfaces address challenges like buckling, payload integration, and environmental adaptability.

Searching arXiv for recent and foundational papers on soft everting robots to ground the article in cited literature. Soft everting robots are compliant robotic systems whose primary motion is generated by eversion: body material turns inside out at a moving interface so that new body length appears at the front, or, in toroidal variants, membrane material is continuously recycled through the body. In the literature this family includes tip-extending or growing robots, everting vine robots, inflated-beam robots, self-propelled toroidal robots, inverting-everting toroidal hydrostats, and ring-shaped helical everting robots. Their defining advantage is that extension or locomotion can occur with very low longitudinal sliding of the exterior surface against the environment, enabling navigation in narrow, cluttered, or fragile spaces while preserving compliance and geometric adaptability (Coad et al., 2019, Perez et al., 2022, 2503.07245).

1. Definition and morphological scope

The most established soft everting architecture is the tip-extending vine robot: a soft tubular body is pressurized internally, and the wall everts at the tip so that the robot extends from its front rather than being pushed from a rigid base. Closely related inflated-beam robots use the same physical principle while emphasizing beam-like mechanics and distributed side actuators. A second major lineage is toroidal: one end everts, another inverts, and the membrane circulates through a closed loop. More recent work expands the family further to underwater hydraulic eversion robots, ring-shaped helical everting robots, and everting toroidal hydrostats used as grippers or prosthetic terminal devices (Coad et al., 2019, Kaleel et al., 2024, Park et al., 3 Mar 2025).

Embodiment	Eversion topology	Representative capability
Tip-extending vine robot	Pressure-driven tip eversion from a fixed base	3-DOF tip position and force control during growth (Coad et al., 2019)
Multi-segment vine robot	Tip eversion with selective pouch actuation	Repeatable multi-turn free-space growth (Kübler et al., 2022)
Underwater hydraulic vine robot	Water-driven eversion with side pouches	68° maximum bending at 2000 mL (Kaleel et al., 2024)
Self-propelled toroidal robot	Simultaneous front eversion and rear inversion	Maze traversal and vertical pipe climbing (Perez et al., 2022)
Helical-ring everting robot	Closed-ring sequential coil eversion	Omnidirectional boundary exploration (2503.07245)
Everting toroidal hydrostat hand	Inverting-everting toroidal hydrostat	Body-powered prosthetic grasping (Park et al., 3 Mar 2025)

This morphological diversity is important because “soft everting robot” does not denote a single kinematic template. Some systems are base-tethered deployers, some are untethered locomotors, and some are manipulators. The common denominator is not a specific geometry but the use of eversion or inversion as the principal means of motion, shape change, or grasp formation.

2. Growth, steering, and reconfigurable shape control

During growth, tip-extending soft robots can control the motion and force of the robot tip in three degrees of freedom using actuators that direct the tip in combination with extension. In the canonical implementation, steering actuators such as cables or pneumatic artificial muscles asymmetrically shorten or lengthen different sides of the body, producing two lateral degrees of freedom plus extension. This establishes the control template that much of the later literature generalizes: eversion supplies longitudinal deployment, while auxiliary actuators impose curvature (Coad et al., 2019).

A common misconception is that adding more side pouches automatically yields arbitrarily complex free-space shapes. When all pouches are serially connected, the whole robot can only perform one constant curvature in free space. The multi-segment selective-steering architecture addresses this by placing pouches in parallel to a pressure supply line and attaching a small magnetic valve to each pouch. A permanent magnet in a motorized tip mount opens only the valve currently passing through the tip, so the newest segment is actuated while earlier segments can remain pressurized and locked in. This system combines 3D-printed magnetic valves, cylindrical pneumatic artificial muscles (cPAMs), a motorized tip mount, and a piecewise constant curvature model. The motorized version lowers the pressure needed for eversion to about 3 kPa versus 6 kPa for the unmotorized version, achieves a maximum measured bending of $0.185^\circ/\mathrm{mm}$, and produced tip-position standard deviations of 11.8 mm for a constant left turn, 4.1 mm for an increasing left turn, and 11.3 mm for an S-turn across five trials; a 2 m long robot also held a manually formed shape for at least 15 minutes after pressure removal (Kübler et al., 2022).

Variable stiffness introduces a different route to shape control. Embedded positive-pressure layer jamming discretizes an everting inflated beam into sections that can be selectively stiffened, so tendon actuation localizes bending at interfaces between stiff and compliant regions. The robot described in this framework is a 55 mm diameter tube made from TPU-coated ripstop nylon, stored inverted on a spool inside a rigid pressure vessel, with three tendons spaced $120^\circ$ apart for 3D bending. Each section contains 6 layer stacks around the circumference, each stack made of 15 layers of copy paper. The measured effective material stiffness increases from 263.2 kPa unjammed to 3038.6 kPa jammed; at 6.9 kPa internal pressure, the force required to reach 10 mm tip deflection increases from 1.07 N unjammed to 6.68 N jammed, a 624% increase, and at 20.7 kPa internal pressure the force increase is 663%. Positive-pressure jamming is nearly as effective as vacuum-based jamming, with differences of 4.7% at 13.8 kPa and 2.1% at 20.7 kPa, which suggests that vacuum hardware is not necessary for eversion-compatible stiffness programming (Do et al., 2023).

Underwater steering demonstrates that the eversion principle transfers to hydraulic operation but not without new constraints. The underwater hydraulic volumetric soft everting vine robot uses three parallel polyethylene tubes: a middle tube for extension and two side pouch sets for bending. Each side contains square volumetric pouches of $10\ \mathrm{cm}\times 10\ \mathrm{cm}$ separated by 3 cm gaps, with 7 pouches per side, 6 of which are used for actuation. Water drives extension, bends the pouches, and contributes to neutral buoyancy in principle. In tests on three prototypes, bending was observed across a range of inflation volumes, with a maximum angle of 68° at 2000 mL, while one image showed 65° at 3228 mL. The general trend was a modest increase from 0 to about 800 mL, a faster increase thereafter, and a plateau around 2400 mL for most robots, though one prototype plateaued earlier around 1600 mL. An important limitation is that the robot must fully extend before steering is possible (Kaleel et al., 2024).

3. Analytical and numerical mechanics

The mechanics of soft everting robots are often organized around force and moment balances rather than rigid-link kinematics. For retraction and inversion, a force-based criterion predicts whether the body will invert or buckle: the robot will do whichever requires the lower tail tension. The inversion tension is modeled as

$T_T = \frac{1}{2}PA + F_I,$

where $P$ is internal pressure, $A=\pi R^2$ is cross-sectional area, and $F_I$ is an offset force associated with material deformation at the tip. Equating inversion force and crushing force yields the minimum pressure for inversion,

$P_{min} = \frac{2F_I}{A},$

which was about 1.1 kPa for the reported robot. For straight robots, inversion occurs only if $T_T$ is lower than both the axial buckling force and the crushing force $F_{crushing}=PA$; because inversion force does not depend on length but buckling force decreases with length, there is a critical length above which a straight robot buckles. For curved robots, the relevant competition is between inversion tension and a transverse-buckling condition with moment arm

$120^\circ$0

which predicts that more curved robots buckle sooner. Experiments matched the model boundary: short robots invert, long robots buckle, and more curved robots buckle at shorter lengths (Coad et al., 2019).

Self-weight imposes a different stability limit in open space. For straight growing inflated beam robots launched at angle $120^\circ$1 above horizontal, the gravitational moment is modeled as $120^\circ$2 with

$120^\circ$3

and the critical folding moment is

$120^\circ$4

For horizontal launch, the special-case collapse length is

$120^\circ$5

The reported scaling laws are that collapse length increases non-linearly with launch angle magnitude, linearly with diameter, and with the square root of internal pressure. A 0.95 m gap-crossing example illustrates the design use of these laws: a robot at $120^\circ$6, 3.24 cm diameter, and 4.14 kPa was predicted to collapse at 0.82 m and failed, whereas increasing launch angle to $120^\circ$7, increasing diameter to 4.04 cm, or increasing pressure to 10.34 kPa enabled success (McFarland et al., 2023).

These analytical models coexist with high-fidelity finite element models because shell inflation, contact, buckling, anisotropy, and large deformation are central rather than secondary effects. A general FEM framework for inflated-beam robots uses Abaqus Explicit with a 1.1 s simulation, 0.1 s for body pressurization and 1.0 s for actuator pressurization, general contact, pinned boundary conditions, shell thicknesses of 200 $120^\circ$8m for the inflated beam and 50 $120^\circ$9m for the actuator, and actuator/body pressures of 10, 20, and 30 kPa over a 2 kPa body baseline. The framework was validated on four actuator types—sPAM/PM, cPAM, ePAM, and fPAM—showing accuracies of 96.4/95.8/94.3% for sPAM/PM, 87.1/97.0/97.3% for cPAM, 80.7/92.1/94.9% for ePAM, and 97.6/97.3/96.7% for fPAM at 10/20/30 kPa. Geometry-based contraction achieved significant deformation with 92.1% accuracy once the buckling pattern formed, but dropped to 80.7% at lower pressures; material-based contraction remained at least 96.7% accurate. This suggests that predictive design for soft everting robots requires both reduced-order mechanics and shell-level numerical modeling, especially when weld-line stress concentrations and anisotropic fabrics are relevant (Pasquier et al., 2023).

4. Retraction, buckling, and controllability

A persistent misunderstanding is that retraction is simply the time-reverse of growth. In practice, pulling the internal body material back from the base often causes the body to buckle rather than invert, especially for long or curved robots. Inversion means the wall flips back through the tip so that the tip retracts in the desired direction; buckling means the wall folds over itself, the tail is pulled inward without inversion, and tip motion becomes laterally unstable and unpredictable. Once buckling begins, the retraction degree of freedom becomes coupled to steering actuation and the environment, and control of both tip position and tip force becomes impossible (Coad et al., 2019).

The anti-buckling solution reported for tip-extending robots is to move the retraction force application point from the base to the tip. A tip-mounted electromechanical retraction device uses motor-driven rollers to grip the tail, routing apertures to guide it, and a grounding ring anchored to the robot tip. Because the pulling force $10\ \mathrm{cm}\times 10\ \mathrm{cm}$0 and grounding force $10\ \mathrm{cm}\times 10\ \mathrm{cm}$1 satisfy

$10\ \mathrm{cm}\times 10\ \mathrm{cm}$2

the effective force-grounding distance for retraction is essentially zero, so the retraction force does not load the deployed body as a long unsupported member. The resulting tail tension becomes

$10\ \mathrm{cm}\times 10\ \mathrm{cm}$3

If the device supplies

$10\ \mathrm{cm}\times 10\ \mathrm{cm}$4

then $10\ \mathrm{cm}\times 10\ \mathrm{cm}$5, and the robot can invert without any tail pulling from the base. This reframes the retraction problem as a grounding problem rather than an intrinsic impossibility of inverse eversion (Coad et al., 2019).

The consequence is a restoration of the same control structure available during growth: steering actuators again determine lateral motion and force, while inversion determines extension or retraction independently. With the device, the robot demonstrated three tasks that had previously been impossible: exploring different branches of a forking path, reversing growth while applying minimal force on the environment, and bringing back environment samples to the base. These demonstrations establish retraction as a controllable motion primitive rather than a failure-prone reset operation (Coad et al., 2019).

5. Tip interfaces, internal transport, and manipulation

Tip interfaces became a central subtopic once soft everting robots began carrying sensors, pouches, and other hardware. Rigid caps and mechanism-heavy tip mounts can fail or jam when body-mounted features protrude from the skin, and they reduce mobility in confined environments. The soft-cap approach replaces those interfaces with a fabric cylinder with a closed rounded end that is slipped onto the tip and retained by distributed friction. Its design is parameterized by robot diameter $10\ \mathrm{cm}\times 10\ \mathrm{cm}$6, cap unstretched diameter $10\ \mathrm{cm}\times 10\ \mathrm{cm}$7, effective friction/contact length $10\ \mathrm{cm}\times 10\ \mathrm{cm}$8, diameter mismatch

$10\ \mathrm{cm}\times 10\ \mathrm{cm}$9

and percentage mismatch

$T_T = \frac{1}{2}PA + F_I,$0

An eight-prototype study showed that performance is governed by a friction tradeoff rather than by retention alone. Prototype 4, a stretch-fabric cap with $T_T = \frac{1}{2}PA + F_I,$1 cm and $T_T = \frac{1}{2}PA + F_I,$2, achieved 100% on multilayer-body eversion, protruding-object eversion, squeezability, and navigability, with 80% payload stability and a 96% overall score. Prototype 7, an elastic-band cap with $T_T = \frac{1}{2}PA + F_I,$3, achieved the highest payload stability among the soft caps at 90% and an overall 98%. By contrast, the tightest designs degraded mobility: prototype 1, with $T_T = \frac{1}{2}PA + F_I,$4 cm and $T_T = \frac{1}{2}PA + F_I,$5, dropped to 30% squeezability and 0% navigability, and prototype 6 achieved only 20% squeezability. This directly contradicts the common assumption that the tightest cap is necessarily the best cap (Suulker et al., 2023).

A related deformable tip mount reformulated the same idea as a textile “beanie” that can remain attached throughout eversion while accommodating protrusions and changing local geometry. In qualitative case studies, a 10 cm diameter robot carrying a camera passed through a 9 cm gate; two protruding pipe connectors of approximately $T_T = \frac{1}{2}PA + F_I,$6 and $T_T = \frac{1}{2}PA + F_I,$7 everted successfully and ended up outside the cap boundary; and navigation layers inflated and reoriented the tip while the cap remained attached. This suggests that textile compliance is not merely a lightweight alternative to rigid caps, but a geometric compatibility mechanism for eversion with layered, protruding, or reconfiguring skins (Suulker et al., 2024).

Internal payload transport extends the idea of the robot as a protected transport channel rather than only a tip-deployer. An origami-inspired soft growing robot with an open internal working chamber, a 10.2 cm inner tube, and a 13 cm overall diameter was analyzed for growth with payloads, payload slip, and unsupported gap crossing. The basic growth condition is

$T_T = \frac{1}{2}PA + F_I,$8

and with payload mass on an incline,

$T_T = \frac{1}{2}PA + F_I,$9

Across experiments on steering, vertical transport, movement through holes, and movement across gaps, the robot transported payloads up to 1.5 kg, moved through discrete turns up to 135°, traversed circular apertures with as little as 0.01 cm clearance around the payload, and crossed unsupported gaps of 1.15 m. The reported friction-loss term from slope fitting was about 3.6 N, close to prior estimates of 4.1 N. A major qualitative conclusion was that transporting the payload after the robot first spans the gap is more effective than carrying it while cantilevered (DeVries et al., 29 Jul 2025).

Manipulation can also be realized with everting bodies. A soft prosthetic hand based on an inverting-everting toroidal hydrostat used a 12.7 cm long, 5.1 cm outer diameter hydrostat inside a flexible housing and was integrated into a body-powered elbow-driven system. Bench experiments found a peak cable tension of 1.6 N for the everting hand, compared with 30.0 N for the Kwawu hand and 28.1 N for the Hosmer hook. Peak pulling force required to remove a grasped object was 15.8 N for the everting hand, versus 6.9 N and 4.0 N. In a pilot user study, the everting hand achieved 19.7 blocks per minute in the Box and Blocks Test, compared with 12.2 and 14.2 for the comparison devices, and users completed several Jebsen-Taylor tasks faster than with the Kwawu hand. The same results also delimit the method: performance was weaker for writing, checker stacking, and large heavy cans, indicating that conformal everting grasping favors adaptable enclosure over fine rigid precision (Park et al., 3 Mar 2025).

6. Locomotion regimes, application domains, and unresolved constraints

Not all soft everting robots are deployers from a fixed base. The self-propelled soft everting toroidal robot uses an air-filled LDPE membrane and an internal propulsion device with two DC motors driving two active rollers, plus passive rollers for stabilization and grounding. As the device retracts the membrane tail, one end of the torus inverts while the other everts, so the outer body remains mostly stationary relative to the environment and membrane material recycles like a three-dimensional tank tread. The robot requires only a single control signal to move, traversed a zigzag maze, squeezed through an 11 cm aperture even though its inflated outer diameter was 13.7 cm, and climbed vertically in about 5 seconds. In a pipe, the coefficient of static friction between LDPE and acrylic was measured as $P$0; about 0.3 kPa internal pressure was sufficient to support its own weight plus battery, and about 80 N support was achieved at 3.45 kPa. The force model also showed that the motor force used to propel the membrane is decoupled from the force used to brace the robot against its environment (Perez et al., 2022).

The WHERE-Bot demonstrates a distinct locomotion regime: a wheel-less, closed-ring, helical everting robot built from a looped Slinky-like spring wrapped around a helical-gear hub. It has a nominal loop diameter of 266 mm, height of 80 mm, mass of 436 g, 37 coils, and a hub helix angle of 5°. Its motion decomposes into spiral-rotation along the hub circumference, self-rotation around the hub center, and orbiting around a point. A steering servo shifts concentrated masses around the ring, and the structural angle $P$1 determines orbit radius, turning direction, travel distance per cycle, and motion period. In a square-boundary experiment, the robot was placed inside a $P$2 boundary, contacted a wall when the orbit radius was large enough, crawled along the wall while rotating, transitioned around the corners, and navigated the whole square boundary after 268.13 s. The design is explicitly intended to exploit morphology and frictional interaction rather than sensor-based boundary detection (2503.07245).

Across these embodiments, the application record spans search-and-rescue, nuclear inspection, archaeology, teleoperation, underwater exploration in coral reefs, crevices, or wreck interiors, sample retrieval, confined-space maze navigation, pipe climbing, hazardous-area supply deployment, minimally invasive or medical-delivery scenarios, and prosthetic grasping (McFarland et al., 2023, Kaleel et al., 2024, DeVries et al., 29 Jul 2025). The range of demonstrated functions indicates that eversion is not only a deployment mechanism but also a platform for locomotion, manipulation, payload transport, and controlled environmental interaction.

Several unresolved constraints recur across the literature. Retraction can remain vulnerable to buckling when force is grounded incorrectly; underwater steering showed poor repeatability across nominally identical prototypes and was affected by leakage, trapped air, water temperature, and the absence of closed-loop control; payload deployment through holes and tight turns introduces membrane wear and model mismatch from pressure-flow effects; the self-propelled toroidal robot is dominated by membrane inversion and roller-transmission losses; the WHERE-Bot is presently limited to hard surfaces with low friction coefficients and exhibits oscillatory motion because of large spring–hub clearance; and the prosthetic toroidal hydrostat raises puncture, visibility, and fine-manipulation issues (Coad et al., 2019, Kaleel et al., 2024, DeVries et al., 29 Jul 2025, Perez et al., 2022, 2503.07245, Park et al., 3 Mar 2025). Taken together, these results suggest that the central research agenda is no longer whether soft everting robots can move, but how to make eversion-compatible steering, retraction, payload integration, stiffness programming, and environment-specific reliability cohere in a single controllable architecture.