Steerable Vine Robots

• Steerable vine robots are soft continuum systems that extend by tip eversion, combining growth and steering to maneuver through confined and complex spaces.
• They employ diverse actuation methods—including pneumatic, tendon-driven, magnetic, hydraulic, and hybrid systems—to achieve precise curvature control and versatile navigation.
• Their design integrates localized stiffness modulation, distributed sensing, and autonomous control strategies to overcome challenges in load-bearing, retraction, and environmental interaction.

Steerable vine robots are soft continuum robots that extend by everting their body material at the tip while actively controlling the direction of growth. The class was formalized as “vine robots” for tip-extending, highly length-changing robots whose movement resembles trailing plants, and it has since diversified into pneumatic, tendon-driven, magnetic, hydraulic, hybrid, preformed, and material-responsive systems for search and rescue, inspection, pipe and burrow traversal, underwater exploration, decontamination, and endoluminal navigation (Coad et al., 2019, Maur et al., 2022). Their defining mechanical feature is that new body material emerges at the distal end while the deployed exterior remains largely stationary relative to the environment, so steering is coupled to growth rather than to wheel slip, jointed backbone translation, or surface dragging (Coad et al., 2019, Kim et al., 2024).

1. Defining mechanics of growth-based locomotion

The canonical vine robot consists of a thin-walled pressurized tube stored in an inverted state at the base and deployed by tip eversion. Early fielded systems used a central growth tube with three outer actuator tubes arranged around it, whereas later systems integrated actuators, sensors, valves, electronics, or steering joints directly at or near the tip (Coad et al., 2019, Maur et al., 2022). In RoBoa, the tube is furled inside a supply box and everted by pressurizing the interior to approximately $15\text{–}20$ kPa gauge; the prototype reported forward speeds up to $6$ m/min in unobstructed corridors and a maximum fully deployed length of $17$ m in trials (Maur et al., 2022). In early field deployments, stored length was $10$ m for the competition robot and $7.5$ m for the archaeology robot, with practical growth speeds of about $6$ cm/s and a reported maximum growth speed of about $10$ cm/s (Coad et al., 2019). Other architectures replace pure pneumatic feed with motorized spools or rollers: PanoVine uses a high-torque brushless DC motor turning a spool at the base, while the real-time two-tape system feeds flattened material externally through rollers that provide both growth and air sealing (Qin et al., 22 Jun 2026, Liu et al., 2 May 2025).

The absence of substantial external sliding is central to the mechanism. RoBoa explicitly reports that the outside of the tube remains stationary relative to rubble, making external friction negligible, while early field systems emphasized that growth lets the tip move without relying on reaction forces from the environment (Maur et al., 2022, Coad et al., 2019). This property also underlies later claims of shear-free navigation in magnetic endoluminal steering and minimal sliding friction during branch selection in pipe and burrow environments (Kim et al., 2024, Qin et al., 9 Jul 2025).

A standard force relation appears in multiple formulations: the tip-driving force scales with internal pressure and cross-sectional area. RoBoa writes the nominal tip force as

$F_{\rm tip}=p_{\rm box}\,\pi R^2,$

and the early field model expresses the maximum extension force as $F_{\max}(P,\Delta L)=P\cdot A-k\cdot\Delta L$, with growth when external resistance remains below that bound (Maur et al., 2022, Coad et al., 2019). These models differ in detail but encode the same trade-off: pressure that enables eversion also sets the available axial load margin for steering hardware, payload, and environmental interaction.

2. Steering architectures and actuation families

The dominant steering family uses asymmetric shortening or lengthening of the body. In the early three-actuator configuration, three series-pouch motors arranged at $\psi_i=0, 2\pi/3, 4\pi/3$ shorten one side of the cylinder and induce curvature toward that side (Coad et al., 2019). Comparative actuator studies then distinguished pouch motors, cylindrical pneumatic artificial muscles (cPAMs), and fabric pneumatic artificial muscles (fPAMs): pouch motors were simplest to prototype, cPAMs produced the highest curvature and force, and fPAMs actuated fastest and everted at the lowest pressure (Kübler et al., 2023). A later steerability study further separated external pouch attachment from integrated pouch construction, showing that externally attached actuators begin curving at low pressure ratio but saturate, whereas integrated actuators require higher pressure ratio before bending but achieve higher curvature overall (McFarland et al., 26 Oct 2025).

A second family concentrates steering authority at the tip. RoBoa places an internal robot immediately behind the head, consisting of a two-segment, three-actuator pneumatic bundle with proportional valves collocated at the chambers and commanded through a single PoE cable. By actuating only the front segment, RoBoa reports bends up to $6$0 and a turning radius down to $6$1 mm, independent of total tube length (Maur et al., 2022). The hybrid continuum-eversion robot likewise places a 2-DOF continuum steering module at the leading end of a pneumatically everting body, with four nylon steering cables routed from stacked servos to the continuum discs (Al-Dubooni et al., 2024). Tendon steering is extended in the reconfigurable pneumatic joint architecture, where four tendons bend selected compliant segments while independently pressurized joints provide localized selective stiffening and shape locking (Oyejide et al., 17 Apr 2026).

A third family achieves multi-segment steering by selective actuation during growth. The multi-segment soft growing robot with selective steering routes each cPAM through its own normally closed magnetic valve; a permanent magnet in the motorized tip mount opens only the frontmost valve as it passes, pressurizing that segment and then trapping the pressure after the valve exits the magnetic field. This enables repeatable growth into different shapes and holding those shapes without environmental contact (Kübler et al., 2022). Related work on preformed vine robots programs discrete bends into the body before deployment using tape, ultrasonic welding, or embedded fastener loops, turning growth into execution of a mechanically encoded path rather than online actuator modulation (Agharese et al., 2023). Real-time two-tape control pursues a similar idea dynamically by autonomously applying adhesive tape to induce surface wrinkles; the reported prototype achieved repeated $6$2 turns in planar motion (Liu et al., 2 May 2025).

A fourth family uses external steering fields or fixtures rather than internal asymmetric contraction. External magnetic actuation mounts an internal permanent magnet at the vine tip and robotically positions an external permanent magnet to generate a tip wrench, yielding a $6$3 mm diameter robot with integrated camera and 6-DOF localization and a minimum bending radius of $6$4 cm at $6$5 kPa (Kim et al., 2024). External tip steering for pipes and burrows instead uses a compact rigid steering device with tendon-driven soft spherical joints and deployable bracing legs, enabling active branch selection in 3D space without distributing actuators along the entire soft body (Qin et al., 9 Jul 2025).

Finally, several architectures depart from standard contracting-side pneumatic steering. The high-curvature inspection robot uses an anisotropic wrinkled composite film whose axial modulus is approximately $6$6 MPa and circumferential modulus approximately $6$7 MPa, with layer jamming inserts acting as brakes so that one side lengthens under body pressure while the other remains locked (Mendoza et al., 2023). Underwater steering uses hydraulic side pouches driven by ambient water, while the light- and heat-seeking system embeds photothermal phase-change series actuators directly in the skin, eliminating wires and centralized control (Kaleel et al., 2024, Deglurkar et al., 2023).

3. Kinematics, mechanics, and simulation frameworks

Constant-curvature and piecewise-constant-curvature assumptions remain the dominant analytical abstractions. In the early teleoperated field robot, actuator pressures are mapped into a planar tip displacement $6$8, then into curvature magnitude $6$9 and bending-plane angle $17$0 under a constant-curvature model (Coad et al., 2019). RoBoa refines this to a pressure-to-curvature mapping for a three-actuator arrangement, with scalar planar bending written as

$17$1

and total deflection over segment length $17$2 written as $17$3 (Maur et al., 2022). The multi-segment selective-steering robot uses a piecewise constant curvature model along the outside of the body and chains segment transforms to predict the final shape, reporting mean tip-path errors of $17$4 mm for a constant left turn and $17$5 mm for an S-turn in high-pressure cases (Kübler et al., 2022).

Alternative steering principles generate different kinematic laws. For plant-inspired asymmetric lengthening, the inspection robot derives

$17$6

so curvature increases directly with local axial strain on the extensible side (Mendoza et al., 2023). For preformed robots, fabrication geometry is converted from desired link lengths, twists, and bend angles into cylindrical fold distances and offsets before deployment (Agharese et al., 2023). For externally steered pipe navigation, the tip steering device is modeled as a prismatic-spherical-spherical chain, while the grown body is treated under a constant-curvature assumption (Qin et al., 9 Jul 2025).

Beyond kinematics, recent work integrates actuation, contact, and growth into unified mechanics. Gao et al. formulate a quasi-static beam balance

$17$7

for a growing pressurized beam under actuation and obstacle contact, then discretize it into rigid links and solve each step by GPU-parallel position-based dynamics with autodiff and a learned actuator surrogate (Gao et al., 18 Sep 2025). This framework is explicitly used for design optimization, including actuator placement and exploitation of environmental contacts to minimize the number of actuators.

Steerability and collapse models supplement these local bending laws. The steerability study reports that steerability decreases with increasing tip load, peaks at moderate chamber pressure rather than at the maximum available pressure, increases with length, and is largely unaffected by diameter once diameter is above collapse threshold (McFarland et al., 26 Oct 2025). The self-weight collapse model starts from the classical moment capacity

$17$8

and generalizes it to arbitrary curved shapes by summing distributed gravity moments, actuator contributions, and tail-tension effects (McFarland et al., 29 Oct 2025). This body of work shifts the literature from qualitative descriptions of “growing around obstacles” toward explicit prediction of when a chosen steering policy, geometry, or payload will remain mechanically feasible.

4. Teleoperation, feedback, localization, and autonomy

The earliest steerable vine robots were designed around human-in-the-loop teleoperation. The field-deployed system of Coad et al. used a custom flexible joystick with an IMU at its tip, two potentiometers, and a direction switch; joystick bending was converted into desired curvature and mapped to the three pneumatic actuators, while separate controls regulated growth pressure and spool motion (Coad et al., 2019). This architecture established the now standard operational split between directional control and growth-rate control.

Later systems pushed low-level control and sensing toward the tip. RoBoa uses a decentralized architecture in which proportional valves ride with the internal robot, reducing the tip connection to a single air hose and single PoE cable. Each valve-terminal microcontroller runs three independent PID loops for chamber pressure, draw-wire sensors provide closed-loop deflection feedback, and the head carries a $17$9 greyscale camera, a 9-axis IMU, and microphone plus speaker, streamed through an onboard SBC (Maur et al., 2022). The selective-steering robot similarly embeds control logic in the tip mount: motorized rollers regulate smooth eversion and retraction, while magnetic valve timing implements a discrete spatial steering sequence without per-segment onboard electronics (Kübler et al., 2022).

Localization and shape estimation have diversified. Magnetic external steering combines an external KUKA-mounted magnet with a flexible magnetic-sensor array around the internal tip magnet, solving the tip pose in real time at $10$0 Hz (Kim et al., 2024). Pipe and burrow navigation fuses a spool encoder with a tip IMU and camera to reconstruct a 3D trajectory onboard and stream position and orientation to the operator (Qin et al., 9 Jul 2025). Distributed IMU shape sensing extends this idea along the whole body: with 18 BNO055 IMUs on a $10$1 m robot, mean tip position error was $10$2 of robot length for passive steering, $10$3 for active steering, and $10$4 across growth experiments from $10$5 cm to $10$6 cm, with an average orientation drift rate of $10$7min across 15 sensors (Laudenslager et al., 27 Feb 2026).

Autonomous control now exists alongside teleoperation. PanoVine distributes 19 USB cameras along a 6 m-long, 0.16 m-diameter everting body and steers using six actively actuated revolute joints. Each image is resized to $10$8, encoded by a CLIP-pretrained ViT-B/16, and fused with proprioceptive features in a Diffusion Transformer with 7 blocks, embedding dimension $10$9, and 8 heads. The policy outputs relative action chunks over horizon $7.5$0 at 5 Hz and is trained from 41 complex-course navigation demonstrations and 80 object-reaching demonstrations (Qin et al., 22 Jun 2026). This marks a methodological break from explicit inverse kinematics: rather than estimating a full state and commanding a geometric model, the controller directly learns closed-loop visuomotor compensation for hysteresis, tether effects, and unexpected deformations.

5. Experimental performance and application domains

Search and rescue has been a recurring benchmark because rubble simultaneously penalizes rigid robots and rewards low-friction growth. Early field work demonstrated navigation over uneven terrain, unstable obstacles, and a small aperture in RoboSoft 2018, as well as exploration of archaeological tunnels containing rocks, horizontal and vertical turns, and a $7.5$1 bend (Coad et al., 2019). RoBoa was then evaluated at the Swiss Rescue Troops’ training site in a 10 m debris tunnel with random rubble, narrow shafts, and $7.5$2 turns; over 5 trials it reported a travel distance to goal of $7.5$3 m, time to locate a victim of 18–25 min with average 20 min, steering repeatability of $7.5$4 segment bend accuracy, and success rate $7.5$5 (Maur et al., 2022). A later urban search and rescue prototype for confined rubble environments added pneumatic muscles for steering and oscillation, an equation-based robot length control plus feedback pressure regulating system for extending and retracting the body, and controlled tests over varying volume ratio, environmental weight, oscillation, and steering; the abstract reports significant penetration depths in cluttered environments and repeated trajectories suitable for mapping and navigating underground paths (Zhou et al., 2024).

Inspection-oriented systems emphasize free-space curvature and interaction force. The high-curvature inspection robot achieved free-space bending radii $7.5$6 mm for a $7.5$7 mm diameter body, supported tip loads up to about $7.5$8 N at $7.5$9 kPa in cantilever configuration, squeezed through a $6$0 mm gap, pushed a $6$1 g mass out of the way, and grew around compound-curvature obstacles (Mendoza et al., 2023). The actuator-comparison study culminated in a 4.8 m cPAM-equipped vine robot that executed a right $6$2 turn, passed under a low bridge, fit through a reduced-diameter $6$3 cm tunnel despite an $6$4 mm nominal diameter, and lifted itself $6$5 cm against gravity (Kübler et al., 2023). Magnetic external steering reported mean bending radii of about $6$6 mm with $6$7 mm over five runs and traversed a 0.5 m tortuous path in 195 s while striking a target in free space (Kim et al., 2024).

Pipe, burrow, and volumetrically constrained environments have motivated specialized steering devices. The externally steered pipe-and-burrow robot reports active branch selection in 3D with maximum steerable angle $6$8, navigation in pipes as small as 2.5 cm radius, average localization error of $6$9 mm over a 1.2 m laboratory trajectory, and a 1.0 m deployment in a California tiger salamander burrow that reconstructed $10$0 m, $10$1 m, $10$2 m and recorded temperature and humidity profiles (Qin et al., 9 Jul 2025). Underwater hydraulic steering achieved a maximum observed bending angle of $10$3 at approximately $10$4 mL in one prototype, with other prototypes reaching comparable maxima of $10$5 at slightly different volumes (Kaleel et al., 2024).

Manipulation and service tasks are beginning to rely on localized stiffness control and hybrid payload delivery. The reconfigurable pneumatic joint robot carried a $10$6 g payload during 1 m free-space eversion, reduced deflection under a 200 g transverse tip load from about 0.28 m in the baseline trunk to about 0.09 m in the reinforced RPJ system, and reported rise time to $10$7 curvature of 0.40 s versus 2.0 s for a compared layer-jamming mechanism (Oyejide et al., 17 Apr 2026). The hybrid continuum-eversion decontamination robot achieved average grid-spray coverage of $10$8 with minimum single-run success $10$9 and maximum $F_{\rm tip}=p_{\rm box}\,\pi R^2,$0 in lab tests inside a glove-box model (Al-Dubooni et al., 2024).

Autonomous whole-body control has now produced quantitative task success in long-horizon deformable navigation. PanoVine achieved $F_{\rm tip}=p_{\rm box}\,\pi R^2,$1 success in 10 randomized complex-course trials containing a $F_{\rm tip}=p_{\rm box}\,\pi R^2,$2 branch, a $F_{\rm tip}=p_{\rm box}\,\pi R^2,$3 slope climb, a 0.3 m unsupported gap, obstacle passes, and a sharp $F_{\rm tip}=p_{\rm box}\,\pi R^2,$4 corner, compared with $F_{\rm tip}=p_{\rm box}\,\pi R^2,$5 for open-loop trajectory replay and $F_{\rm tip}=p_{\rm box}\,\pi R^2,$6 for a no-rebalancing policy. In object reaching, it reported $F_{\rm tip}=p_{\rm box}\,\pi R^2,$7 success over 20 randomized trials, whereas a single-camera policy achieved $F_{\rm tip}=p_{\rm box}\,\pi R^2,$8 (Qin et al., 22 Jun 2026).

6. Constraints, misconceptions, and emerging directions

Several recurring misconceptions are contradicted by the literature. One is that vine robots steer only by continuously pressurizing symmetric side actuators. In fact, steering may be realized by tip-mounted pneumatic bundles, tendon-driven soft joints, selective magnetic valves, external magnetic wrenches, preformed folds, hydraulic pouches, adhesive-tape-induced wrinkles, or embedded photothermal actuation (Maur et al., 2022, Qin et al., 9 Jul 2025, Kim et al., 2024, Agharese et al., 2023, Kaleel et al., 2024, Deglurkar et al., 2023). Another is that environmental contact is either always necessary or always undesirable. The multi-segment selective-steering robot was explicitly designed to navigate complex paths without interacting with the environment, whereas the contact-aware simulation framework deliberately exploits obstacle contact to reduce actuator count and improve design efficiency (Kübler et al., 2022, Gao et al., 18 Sep 2025).

Retraction remains an unresolved boundary condition rather than a solved inverse of growth. Early field work reported no reliable active retraction because growth reversal caused buckling rather than inversion, and RoBoa noted that inward everting led to kinking in the present design (Coad et al., 2019, Maur et al., 2022). Magnetic steering showed preliminary smooth retraction only when an external magnetic wrench stabilized the tip, and even then occasional tip engulfment occurred under inflation because of sealing-ring friction (Kim et al., 2024). By contrast, the reconfigurable pneumatic joint architecture demonstrated cascading retraction, indicating that localized stiffening may be one route to practical shortening without gross collapse (Oyejide et al., 17 Apr 2026).

A second major limitation is the compliance–stiffness trade-off. Free-space growth, payload carrying, and gap crossing are constrained by low axial stiffness and self-weight collapse. The collapse model of steered robots under self-weight predicts failure when the current gravity moment exceeds a pressure-, tail-tension-, and actuator-dependent collapse moment, and experiments showed that modest actuator inflation can be sufficient to prevent collapse in gap crossing (McFarland et al., 29 Oct 2025). The steerability study similarly found that higher chamber pressure is not monotonically beneficial; steerability peaked at moderate pressure, decreased with tip load, and depended strongly on actuator integration strategy (McFarland et al., 26 Oct 2025). The RPJ literature frames the same issue as a need to decouple global compliance needed for eversion from localized rigidity needed for load-bearing and shape retention (Oyejide et al., 17 Apr 2026).

Sensing and autonomy remain limited by drift, occlusion, and data demand. Distributed IMU shape sensing is feasible but accumulates error with time and length, and active steering increases mean tip error relative to passive steering (Laudenslager et al., 27 Feb 2026). Vision-based autonomy avoids explicit modeling of hysteresis and tether interactions, but present systems still rely on narrow-FOV RGB cameras, lack direct force feedback, and require tens of demonstrations per task family (Qin et al., 22 Jun 2026). Manufacturing and scaling issues are equally persistent: long-vine selective steering exposed valve alignment, leak rates, and tolerance sensitivity; the high-curvature inspection robot remained a proof-of-concept limited to approximately 30 cm length; and underwater hydraulic steering showed prototype-dependent volume-to-angle behavior due to leaks, trapped air, and fabrication variability (Kübler et al., 2022, Mendoza et al., 2023, Kaleel et al., 2024).

Taken together, the literature suggests that the next phase of steerable vine robot research will depend less on any single steering mechanism than on co-design across body material anisotropy, localized stiffness modulation, distributed sensing, and closed-loop control. That interpretation is consistent with current efforts that combine explicit mechanics, contact-aware simulation, field teleoperation, and end-to-end visuomotor learning rather than treating growth, steering, sensing, and deployment as separable subsystems (Gao et al., 18 Sep 2025, Qin et al., 22 Jun 2026).