Crawl N' Sense: Sensing-Based Robotic Locomotion

• Crawl N’ Sense is a robotic locomotion strategy that couples crawling with in-situ sensing, using force–depth data to map terrain properties.
• It employs a dedicated penetration phase where one leg acts as a penetrometer while the others form a stable support, ensuring quasi-static force estimation.
• This method enhances terrain characterization by accurately distinguishing compaction levels and brittle crust ruptures, outperforming faster, inertial gaits.

Crawl N’ Sense denotes a sensing-oriented mode of robotic locomotion in which crawling is used not only for displacement but also as a structured measurement process. In its explicit form, the term names a quadrupedal gait that turns one leg into a penetrometer while the other three legs maintain a stable support triangle, allowing terrain strength and brittle surface-crust rupture to be measured during locomotion (Fulcher et al., 26 Sep 2025). More broadly, related work in soft robotics, snake robotics, and crawl-space locomotion uses the same coupling of body deformation, frictional interaction, proprioception, exteroception, and adaptive control: crawling is shaped by what the robot senses, and sensing quality depends on how the robot crawls (Tirado et al., 5 Jun 2025, Arbelaiz et al., 2024, Chang et al., 2019, Ma et al., 13 Aug 2025).

1. Named gait and core operating principle

In the quadrupedal formulation, Crawl N’ Sense is a sensing-oriented gait implemented on a Ghost Robotics Spirit 40 quadruped with mass $12\,\text{kg}$, body length $0.55\,\text{m}$, three degrees of freedom per leg, and quasi-direct-drive actuation with low gear ratios of $6{:}1$ at the abduction and hip joints and $12{:}1$ at the knee (Fulcher et al., 26 Sep 2025). The low gear ratio is used because it makes the actuators more transparent, so measured motor torques are more useful for estimating external ground reaction forces. Joint position, velocity, and torque are logged at $1\,\text{kHz}$, while IMU-based body pose estimation is combined with kinematics to estimate toe position, velocity, and forces.

The gait has two phases per step. In the penetration phase, three legs support the body in a stable tripod and the fourth leg moves downward at constant speed into the soil, with joint position and current used to compute ground reaction force. In the transition or recirculation phase, the robot shifts its center of mass forward onto the next support triangle and prepares the next leg for penetration. The reported parameters are a stride frequency of $0.05\,\text{Hz}$, step length of $10.5\,\text{cm}$, and penetration speed of $8.0\,\text{cm/s}$ (Fulcher et al., 26 Sep 2025). The body remains still during single-leg sensing, which is explicitly intended to minimize inertial effects and improve measurement resolution.

The principal comparison in the same study is with Trot-Walk, a locomotion-oriented baseline gait from the Ghost Robotics SDK. Trot-Walk uses alternating diagonal-leg pairs, is manually controlled around $2\,\text{Hz}$ with an actual range of $1\!-\!4\,\text{Hz}$, and yields contact or penetration durations of only $0.55\,\text{m}$0 (Fulcher et al., 26 Sep 2025). Because it has no dedicated penetration phase, its force measurements mix ground reaction and leg inertia, making sensing noisier.

Gait	Key kinematic parameters	Sensing consequence
Crawl N’ Sense	$0.55\,\text{m}$1, $0.55\,\text{m}$2, $0.55\,\text{m}$3	Dedicated penetration interval
Trot-Walk	around $0.55\,\text{m}$4, actual $0.55\,\text{m}$5, $0.55\,\text{m}$6 contact	Impact-dominated, noisier force inference

This contrast is central to the concept. Crawl N’ Sense is not merely a slow gait; it is a gait in which locomotion is deliberately restructured so that force–depth data become measurable in a quasi-static regime.

2. Terramechanical inference and rupture detection

The terramechanical objective is to estimate penetration resistance, distinguish homogeneous sand from layered terrain, and detect brittle crust rupture while the robot moves (Fulcher et al., 26 Sep 2025). For homogeneous sand, the force response is treated as approximately linear in penetration depth during slow intrusion, and the paper defines penetration resistance $0.55\,\text{m}$7 as the slope of a linear regression over the active penetration interval: $0.55\,\text{m}$8 For Crawl N’ Sense, the penetration interval is selected with force thresholds of $0.55\,\text{m}$9, which exclude recirculation artifacts and define the quasi-static penetration window. For Trot-Walk, the dynamic contact makes threshold-based interval detection unreliable, so the interval is manually labeled in post-processing.

Ground-plane estimation is gait-dependent. For Crawl N’ Sense, the plane is fit from the three supporting toe contacts: $6{:}1$0 For Trot-Walk, because two legs are recirculating, the same three-point support-plane method is not directly available; past toe locations are propagated using numerical integration of the robot dynamics, and a regression-based method uses current and past toe locations to estimate the ground plane. On deformable terrain, the support toes sink, so the plane estimate is corrected by using the penetrating toe’s position at initial contact as the origin of the depth frame.

The reported laboratory strength measurements show clear terrain separation. Ground-truth penetrometer values were $6{:}1$1 for medium compaction, $6{:}1$2 for low compaction, and $6{:}1$3 for high compaction (Fulcher et al., 26 Sep 2025). Crawl N’ Sense closely tracked ground truth in all three compaction zones, successfully distinguished low, medium, and high compaction, and showed better consistency between left and right forelimb estimates. Trot-Walk could qualitatively distinguish low versus high compaction, but consistently overestimated strength and had larger variance, especially in medium compaction sand.

For crust rupture, the raw force signal at $6{:}1$4 is filtered with a 4th-order Savitzky–Golay filter with a $6{:}1$5 window, and a rupture is flagged when the force drop exceeds $6{:}1$6 (Fulcher et al., 26 Sep 2025). The measured performance sharply separates the two gait classes. Crawl N’ Sense achieved specificity $6{:}1$7, with 3 false positives out of 73 rupture-less steps, and sensitivity $6{:}1$8, missing 4 out of 11 rupture events. Trot-Walk achieved specificity $6{:}1$9, with 56 false positives in 64 rupture-less steps, and sensitivity $12{:}1$0; the paper explicitly interprets this as misleading because the gait essentially flagged almost every step as a rupture. The slower crawl gait can therefore detect brittle ruptures of surface crusts with significantly higher accuracy than the faster trot gait (Fulcher et al., 26 Sep 2025).

The motivating field context is a lunar-analogue transect at Mt. Hood, Oregon, where six sampling locations over a $12{:}1$1 transect showed strength rising from about $12{:}1$2 to $12{:}1$3, with brittle failure from a thin icy crust at some locations (Fulcher et al., 26 Sep 2025). This suggests that dense, step-by-step sensing is not an aesthetic design choice but a response to short-range spatial heterogeneity.

3. Mechanical prerequisites: friction control, anchoring, and peristaltic coordination

The broader crawl-and-sense literature shows that sensing during crawling depends on locomotion mechanics that can create reliable anchors and controlled slip. In the earthworm-inspired soft crawler, locomotion is generated by actively switching friction at the ends while a central pneumatic actuator produces peristaltic body motion (Ge et al., 2017). Inflated end actuators expose silicone to the ground and create high friction; deflated end actuators retract into a hard smooth casing and create low friction. The friction force is modeled as

$12{:}1$4

while the central actuator force is

$12{:}1$5

A central theoretical result of that work is that no crawling is possible without friction (Ge et al., 2017). In the frictionless reduced model,

$12{:}1$6

the controllability matrix has rank 2, the center of mass satisfies

$12{:}1$7

for all reachable states from the origin, and the system cannot translate. In the idealized friction-enabled model with augmented inputs $12{:}1$8, the controllability matrix has rank 4, which the paper uses to show that friction is the enabling mechanism for locomotion. This point is important in Crawl N’ Sense because the quality of force-based sensing depends on whether the body can establish a stable anchor while another segment probes.

The same anchoring logic appears in the multimodal limbless soft robot with a kirigami skin (Tirado et al., 5 Jun 2025). That robot uses two antagonistic pairs of fiber-reinforced inflatable actuators, a foldable kirigami skin, a sensorized head with two proximity sensors, and a pointy tail. The actuated link lengths are modeled as

$12{:}1$9

$1\,\text{kHz}$0

and the paper reports that $1\,\text{kHz}$1 is the best phase shift for rectilinear locomotion, whereas $1\,\text{kHz}$2 reduces anchoring efficiency and lowers propulsion. The kirigami skin is designed so that friction is lower in the forward direction and higher in the backward direction, and experimental friction tests show that in most cases backward friction exceeds forward friction (Tirado et al., 5 Jun 2025). The paper defines the friction coefficient as

$1\,\text{kHz}$3

and interprets the resulting traction as a two-anchor effect: one segment grips while the other moves.

Performance data reinforce the point. For rectilinear locomotion, the best performance was observed at $1\,\text{kHz}$4 to $1\,\text{kHz}$5 and $1\,\text{kHz}$6, with peak speeds up to about $1\,\text{kHz}$7 on a fine surface and up to about $1\,\text{kHz}$8 on a coarse surface (Tirado et al., 5 Jun 2025). The coarse substrate improved traction because the larger pores gave better grip. In both the earthworm and kirigami systems, sensing is not separable from traction design; controlled friction is what makes the body’s force signatures interpretable.

4. Exteroceptive crawl-and-sense loops

A second strand of the literature treats Crawl N’ Sense as a closed loop between locomotion and exteroception. In the multimodal limbless robot, the “sense” component is implemented with two time-of-flight proximity sensors mounted on the head, each with range $1\,\text{kHz}$9, field of view $0.05\,\text{Hz}$0, and mounting offset $0.05\,\text{Hz}$1 for lateral awareness (Tirado et al., 5 Jun 2025). Sensor data are sent via I2C to an ESP32-C3 and forwarded over WiFi to the high-level controller or HMI. The paper defines

$0.05\,\text{Hz}$2

from right and left sensor readings and uses explicit obstacle logic: if obstacles are farther than $0.05\,\text{Hz}$3, the robot proceeds straight; if obstacle distance is between $0.05\,\text{Hz}$4 and $0.05\,\text{Hz}$5, steering guidance is generated from $0.05\,\text{Hz}$6; if an obstacle is within $0.05\,\text{Hz}$7, the system overrides user command and forces evasive turning. In a cluttered arena with a coarse foam substrate and three obstacles, the robot used rectilinear crawling with $0.05\,\text{Hz}$8 and $0.05\,\text{Hz}$9, steering gaits, and real-time proximity feedback, and successfully reached the goal in 18 minutes (Tirado et al., 5 Jun 2025).

In the snake-robot literature, the sensing loop is monocular and vision-based rather than proximity-based (Chang et al., 2019). A 12-link snake robot with scales carries a wireless analog color camera on its head, streams $10.5\,\text{cm}$0 images to an external PC, and uses ORB-SLAM for self-localization in unknown planar obstacle fields. The rectilinear gait is reduced at steady state to a fixed-speed unicycle-like model, with steering control represented by body curvature $10.5\,\text{cm}$1 and turning rate $10.5\,\text{cm}$2 given as a linear function of $10.5\,\text{cm}$3. The inverse mapping is written

$10.5\,\text{cm}$4

The perception pipeline combines ORB-SLAM, ground-plane traversability segmentation, perception-space collision checking, and receding-horizon planning (Chang et al., 2019). The traversability classifier is a DCT-based SVM that labels $10.5\,\text{cm}$5 pixel blocks as ground or non-ground. Candidate trajectories are collision-checked by projecting a synthetic binary head footprint into the segmented image; the planner evaluates $10.5\,\text{cm}$6 candidate trajectories, including one straight path and symmetric left- and right-curving paths, replans every 10 seconds or about 4 gait cycles, and tracks the selected path by choosing the furthest waypoint within a radius $10.5\,\text{cm}$7. The system was demonstrated on a carpeted surface in a cluttered indoor obstacle field, where it navigated around obstacles, performed multiple replans, traversed the field, and exited into a corridor (Chang et al., 2019).

These exteroceptive systems show that crawl-and-sense need not mean terrain probing alone. It can also mean that gait selection, steering, and collision avoidance are continuously reconfigured from onboard perception while crawling proceeds.

5. Proprioception, endogenous control, and learned adaptive crawling

Another major interpretation of Crawl N’ Sense is that the robot senses its own body and uses that information to generate or adapt the crawl itself. In “Excitable crawling,” a single soft crawler segment is controlled by a FitzHugh–Nagumo-inspired bistable electrical controller that receives proprioceptive strain feedback (Arbelaiz et al., 2024). The body strain is

$10.5\,\text{cm}$8

the actuation force is

$10.5\,\text{cm}$9

and the proprioceptive current is

$8.0\,\text{cm/s}$0

With $8.0\,\text{cm/s}$1, the voltage dynamics become

$8.0\,\text{cm/s}$2

The paper’s central claim is that the controller is bistable on its own, but with strain feedback it repeatedly crosses switching thresholds and emits spikes, producing endogenous crawling rather than an externally programmed periodic gait (Arbelaiz et al., 2024).

The reinforcement-learning literature reaches a related conclusion by different means. In the segmented soft crawler modeled after D. melanogaster larvae, locomotion emerges from a closed brain-body-environment loop in which proprioceptive sensing identifies the index of the most strongly contracted segment,

$8.0\,\text{cm/s}$3

and Q-learning learns which neuron to activate (Mishra et al., 2020). Starting from initial all-to-all coupling, learning converges to a sparse nearest-neighbor-like wiring that produces a localized wave of contraction from tail to head. With regularization $8.0\,\text{cm/s}$4, the gait is smoother, more biologically plausible, and more robust to proprioceptive noise; with $8.0\,\text{cm/s}$5, it is about 10% faster but less robust (Mishra et al., 2020). The paper therefore frames crawling as a speed–robustness tradeoff under noisy sensing rather than as a fixed central pattern.

Model-based reinforcement learning extends this logic to onboard noisy sensors (Gzenda et al., 7 Oct 2025). In the 1D inchworm-like soft robotic crawler, the observation stream is

$8.0\,\text{cm/s}$6

where IMUs mounted on the head and base provide acceleration readings and TOF sensors estimate positions relative to a reference object. The latent state is

$8.0\,\text{cm/s}$7

and a Dreamer-style actor-critic pipeline learns latent dynamics for short-horizon prediction while optimizing a periodic gait parameterization

$8.0\,\text{cm/s}$8

In simulation, the task is to move 50 cm forward within a 200 s episode, and the rollout figure shows that the crawler reaches the target in about 14 seconds (Gzenda et al., 7 Oct 2025).

A separate line of work studies whether control should be distributed or centralized when sensing and actuation are both rudimentary (Gagliardi et al., 3 Jun 2025). In the binary sucker-and-spring crawler, each local controller observes only elongation versus compression and chooses adhere or release. Purely distributed learning is sufficient for effective crawling, but centralizing control enhances speed and robustness to failure. For $8.0\,\text{cm/s}$9, the fully centralized controller improved performance by about 17% over distributed standard control and about 30% over distributed hive control; however, the Q-table for a single control center at that size has about $2\,\text{Hz}$0 entries, whereas two control centers reduce this to about 2000 entries (Gagliardi et al., 3 Jun 2025). The computational cost therefore scales exponentially with the number of suckers per controller. This suggests that crawl-and-sense architectures are shaped not only by body mechanics and sensors but also by the structure of the control substrate that interprets those signals.

6. Applications, scope, and limitations

The application domains attached to Crawl N’ Sense are consistently those in which mobility and measurement must coexist. The sensing-oriented quadrupedal gait is motivated by geological field science and planetary exploration, where the robot is treated as a moving proprioceptive penetrometer capable of dense geotechnical mapping rather than sparse stop-and-probe sampling (Fulcher et al., 26 Sep 2025). The soft limbless multimodal crawler is positioned for search and rescue, environmental monitoring, and industrial inspection in confined or unstructured environments (Tirado et al., 5 Jun 2025). The crawl-space locomotion literature targets low-ceiling tunnels, caves, stair tunnels, and similar spatially constrained environments in which exteroceptive sensors are unreliable because of low illumination, smoke, occlusion, or tight geometry (Ma et al., 13 Aug 2025).

The point-cloud-supervised proprioceptive locomotion framework PPL is especially explicit about this operational setting (Ma et al., 13 Aug 2025). At run time, the actor uses only proprioception and learned state estimates, while point cloud information appears only as training-time supervision. The policy input includes

$2\,\text{Hz}$1

together with estimated base velocity, collision states for head, base, hip, foot, thigh, and shank, and a latent ground/spatial representation $2\,\text{Hz}$2. Real-robot tests on the Unitree Go2 reported 90% success for forward and backward traversal of a $2\,\text{Hz}$3 flat tunnel, 70% success for lateral traversal, and 90% success for both step-up and step-down in a $2\,\text{Hz}$4 stairs tunnel with $2\,\text{Hz}$5 steps and $2\,\text{Hz}$6 step width (Ma et al., 13 Aug 2025). The paper also states that the method is robust to darkness and smoke-filled conditions.

Several limitations recur across the literature. The sensing-oriented crawl is slower than locomotion-oriented baselines and sacrifices coverage rate for measurement fidelity (Fulcher et al., 26 Sep 2025). The multimodal limbless crawler uses assisted teleoperation rather than full autonomy (Tirado et al., 5 Jun 2025). The monocular snake-navigation framework relies on a flat ground assumption and samples the camera only once per gait cycle, about every 2.5 seconds (Chang et al., 2019). The latent-dynamics MB-RL crawler is validated only in simulation and uses a simplified 1D platform (Gzenda et al., 7 Oct 2025). Centralized learning improves speed and robustness, but becomes untreatable with crawler size because of exponential state–action growth (Gagliardi et al., 3 Jun 2025).

A common misconception is that sensing can be appended to an already optimized crawling gait with little effect on locomotion. The literature argues otherwise, often implicitly. The quadrupedal results show that gait choice changes force magnitude, variance, depth estimation quality, and rupture detectability (Fulcher et al., 26 Sep 2025). The friction-controlled and kirigami crawlers show that anchoring mechanics determine whether body deformation produces interpretable translation at all (Ge et al., 2017, Tirado et al., 5 Jun 2025). The proprioceptive and learning-based studies show that the controller may need to be co-designed with the sensing loop, or even generated by it (Arbelaiz et al., 2024, Mishra et al., 2020). Taken together, these works suggest that Crawl N’ Sense is best understood as a co-design principle: locomotion is organized to improve sensing, and sensing is organized to stabilize or adapt locomotion.