MetaRoundWorm: Worm Locomotion & Applications

• MetaRoundWorm is a multifaceted research term denoting varied studies on worm locomotion, behavior, and drug discovery with distinct experimental setups.
• Researchers employ mechanical models, output-driven feedback systems, and unsupervised learning to analyze C. elegans motion and develop bioinspired robotics.
• The term underscores a non-standardized, heterogeneous approach to measuring worm-like motion across different platforms and methodologies.

MetaRoundWorm is not a standardized single designation in the arXiv material considered here. In the supplied corpus, the label is used for multiple, non-equivalent objects associated with roundworm- or earthworm-related research: a study of C. elegans locomotion in structured environments, a feedback system control workflow for discovering a four-drug anthelmintic combination in C. elegans, an earthworm-inspired underwater multi-mode locomotion robot, and a spatio-temporal framework for behavioural classification from single-point worm tracking [(Majmudar et al., 2010); (Ding et al., 2022); (Fang et al., 2021); (Antonic et al., 30 Sep 2025)]. The term therefore indexes a heterogeneous set of problems spanning low-Reynolds-number locomotion, microfluidic behavioural phenotyping, multi-mode bioinspired robotics, and unsupervised movement analysis.

1. Terminological scope and usage

Across the cited works, MetaRoundWorm does not denote a single canonical method, platform, or organismal theory. Instead, it appears as a label attached to distinct research programs whose only commonality is their relation to nematode- or earthworm-like locomotion, behaviour, or control [(Majmudar et al., 2010); (Ding et al., 2022); (Fang et al., 2021); (Antonic et al., 30 Sep 2025)].

Usage	Paper	Core referent
Structured locomotion	(Majmudar et al., 2010)	C. elegans in micro-pillar arrays and shrinking fluid layers
Drug-combination optimization	(Ding et al., 2022)	FSC-guided discovery of a four-drug anthelmintic cocktail
Bioinspired robotics	(Fang et al., 2021)	Earthworm-inspired underwater multi-mode robot
Behavioural analysis	(Antonic et al., 30 Sep 2025)	Single-point spatio-temporal classification of C. elegans locomotion

A recurrent misconception would be to treat MetaRoundWorm as a formal family name. One source is explicit that the robot paper “does not describe ‘MetaRoundWorm’ as a formal family name or a platform series,” and that, if used as a label, it would be an informal name for that particular segmented earthworm-like underwater robot rather than an established family of robots (Fang et al., 2021). The broader corpus reinforces that non-unified usage.

2. Structured-environment locomotion in C. elegans

In the locomotion study, the central object is undulatory motion of C. elegans in environments where hydrodynamic interactions and geometrical constraints are both operative. The primary experiment places worms in a square lattice of PDMS micro-pillars filled with aqueous buffer. The worm length is about 1 mm, the worm thickness about 60 μm, pillar dimensions are 300 μm wide and 300 μm tall, and pillar center-to-center spacing ranges from 380 μm to 610 μm. The typical free-swimming speed is $U \sim 0.1$ mm/s, yielding a low-Reynolds-number estimate

$Re = \frac{UL}{\nu} \sim 0.1$

so viscous forces dominate inertia (Majmudar et al., 2010).

Within the pillar lattice, the worms display a rich set of trajectories. In diagonal locomotion, the worm stays in contact with pillars and pushes against them, producing enhanced locomotion along diagonal paths. The reported diagonal velocity is

$V_{\text{diag}} = 0.8 \text{ mm/s}$

which is about an order of magnitude larger than free swimming speed. In other configurations, the worm follows closed circular or looping paths around the microstructure and may eventually leave one loop and enter another. The pillars thus act simultaneously as mechanical supports for pushing and as geometric guides that constrain the body into preferred paths.

The associated simulation uses a self-locomoting chain of beads moving through a lattice of circular obstacles. The model is described as purely mechanical, with no sensing or behavior, and includes hydrodynamics, contact interactions, geometrical constraints, and an imposed torque wave along the bead chain. This setup reproduces enhanced diagonal motion, constant contact with obstacles, gait-like undulatory motion, and closed circular trajectories. In the periodic-orbit case, the model bead-chain can become locked indefinitely, whereas the real worm can transition from one orbit to another. The stated interpretation is that real worms possess additional variability or biological flexibility not included in the purely mechanical model.

The same paper also documents a swimming-to-crawling transition in a structureless fluid environment: many nematodes are placed in a drop of buffer on an agar plate, the fluid gradually wicks into the agar, the worms collect together, and locomotion shifts from swimming to crawling as fluid volume decreases and confinement increases. The paper frames these results as evidence that undulatory locomotion in complex media is governed by the interplay of body wave kinematics, fluid viscosity, obstacle geometry, and contact mechanics.

3. Output-driven feedback system control for anthelmintic combinations

In the drug-discovery usage, MetaRoundWorm denotes a feedback system control strategy for identifying an effective four-drug anthelmintic combination against wild-type Caenorhabditis elegans. The biological motivation is resistance: parasitic nematodes contribute to major disease burden in humans and livestock, while the limited number of available anthelmintics is losing efficacy as multidrug resistance evolves. The proposed response is not first-principles drug design, but an engineering optimization loop that searches combinations of existing compounds in an output-driven, iterative manner (Ding et al., 2022).

The experimental system places single freely swimming L4-stage worms into a microfluidic drug environment containing three capsule-shaped chambers. Chambers are filled with either M9 buffer or drug solutions. Imaging is performed for 600 s at 1 frame/s, and a custom C++ worm-tracking program extracts the body from the background and records centroid coordinates. Two behavioural outputs drive the optimization. Centroid velocity is the average speed of the worm’s body centroid normalized to control and is the primary measure of reduced movement or paralysis. Track curvature is computed from the path using a Menger curvature formulation; active worms move in relatively straight tracks, whereas stressed or paralyzed worms struggle in place or wiggle erratically, increasing curvature.

The paper distinguishes three movement states in representative data: active with velocity ≈ 134.44 µm/s and curvature ≈ 9.13 mm$^{-1}$, semiactive with velocity ≈ 99.21 µm/s and curvature ≈ 15.84 mm$^{-1}$, and inactive with velocity ≈ 5.5 µm/s and curvature ≈ 56.66 mm$^{-1}$. The winning combination produced minimal centroid velocity (<10 µm/s) and high track curvature (>50 mm$^{-1}$), and the effect appeared within the first minute, which the authors note could potentially allow shortening the assay to around 2 min.

Before optimization, dose responses were measured for four anthelmintics:

Drug	EC50	Winning concentration
levamisole	2.23 µM	1 µM
pyrantel	107.5 µM	12.5 µM
tribendimidine	4.68 µM	4 µM
methyridine	1672 µM	100 µM

The FSC framework is described as model-less and output-driven. It does not require a mechanistic model of worm pharmacology, signaling pathways, or animal biology. Instead, it treats the system as a black box mapping drug concentrations to worm behaviour and iteratively searches for inputs that minimize movement, i.e., low centroid velocity and high track curvature. The search uses differential evolutionary search with six concentration keys for each drug; the highest key corresponds to the EC50 value, and lower keys represent progressively smaller concentrations, typically about half steps. Candidate combinations are divided into a retained P group and a trial T group. The workflow begins with 8 combinations in iteration 1, keeps the best 4 combinations as the new P group, generates 4 new T combinations by differential evolution, and repeats for four iterations.

The final winning cocktail was found in the fourth iteration, specifically T3: levamisole 1 µM, tribendimidine 4 µM, pyrantel 12.5 µM, and methyridine 100 µM. The paper states that this four-drug combination was more potent than any one of the four drugs alone, even though each drug was kept below its EC50 concentration. A plausible implication is that the framework is intended to exploit combination effects without requiring explicit pharmacological synergy models.

4. Earthworm-inspired underwater multi-mode locomotion robot

In the robotics paper, MetaRoundWorm refers to a novel earthworm-inspired underwater multi-mode robot designed for underwater pipeline inspection and underwater resource exploration. The system is a single prototype that combines two locomotion capabilities in one body: earthworm-like in-pipe crawling and propeller-based three-dimensional swimming. Its biological inspiration is explicitly drawn from the metameric segmented body of earthworms, their peristaltic locomotion, antagonistic muscle action, setae-like anisotropic friction, and retrograde peristaltic wave propagation (Fang et al., 2021).

The architecture alternates a peristalsis module and a swimming module to distribute mass more uniformly. The crawling subsystem uses six peristalsis segments, servo motors, cords, pre-bent spring-steel belts, and silicone rubber skins. A single segment includes acrylic plates, eight spring-steel belts, a servomotor, a servo-motor-driven cord, and silicone skin covering. When the cord is pulled, the segment shortens axially and expands radially, while the spring-steel belts restore shape and emulate circular muscle behavior. The swimming subsystem uses four independently controlled propellers: two for horizontal-plane motion and two for vertical-plane motion. A waterproof servo motor and parallel four-bar linkages drive a contract/expand mechanism for the left and right propeller assemblies, allowing retraction for pipe entry and expansion for swimming efficiency. The head and tail include cusps to reduce cross-sectional resistance.

In pipe mode, locomotion is generated by anchoring against the pipe wall and coordinating contraction and relaxation to produce a retrograde peristaltic wave. The reported gait is $n_A = 1$, $n_R = 1$, $k = 1$, with actuation angle 145°. The experiment used a pipe of length 1.8 m and inner diameter 129 mm; the gait period was 6 s; and the robot traveled 0.428 m in 10 periods, giving an average speed of 7.13 mm/s. The crawling model is

$Re = \frac{UL}{\nu} \sim 0.1$0

with $Re = \frac{UL}{\nu} \sim 0.1$1 the total number of robot segments, $Re = \frac{UL}{\nu} \sim 0.1$2 the number of driving modules, $Re = \frac{UL}{\nu} \sim 0.1$3 the number of anchoring segments, $Re = \frac{UL}{\nu} \sim 0.1$4 the number of relaxing/contracting segments, $Re = \frac{UL}{\nu} \sim 0.1$5 the axial deformation of one segment, $Re = \frac{UL}{\nu} \sim 0.1$6 the time of one actuation, and $Re = \frac{UL}{\nu} \sim 0.1$7 the coefficient term used in the cited model. For the six-segment robot, the reported maximum average speed is 2.77 cm/s and the minimum average speed is 0.46 cm/s, depending on gait.

In swimming mode, the robot can move forward and backward, sink and rise, and perform planar turning or circular motion. The equivalent dynamic model treats the robot as a spheroid. The main assumptions are that buoyancy equals gravity and their centers coincide, that the robot is approximated as a spheroid, and that the mass center is at the origin of the local coordinate system. Representative thrust relations are

$Re = \frac{UL}{\nu} \sim 0.1$8

and, in the vertical plane,

$Re = \frac{UL}{\nu} \sim 0.1$9

CFD validation compared circular, forward, and sinking motions using CAD-based simulation and numerical solution with ODE45. For circular motion, the CFD diameter was 1.756 m and the ODE45 diameter 1.733 m; the CFD steady-state speed was 0.855 m/s and the ODE45 steady-state speed 1.079 m/s. For forward motion, the CFD speed was 4.61 m/s and the ODE45 speed 0.374 m/s. For sinking motion, the CFD speed was 0.251 m/s and the ODE45 speed 0.234 m/s.

Experimental validation used a Bestway swimming pool of $V_{\text{diag}} = 0.8 \text{ mm/s}$0 m, water depth 0.4 m, three GoPro HERO8 Black 4K waterproof cameras, a DC 7.0 V, 5.5 A power supply, and an Arduino-based control system. Reported observations include forward travel of about 1.15 m, sink/rise deviation of about 0.077 m, and a circular trajectory fit with a radius around 2.98 m in one test. A key practical result is that the robot can crawl out of a pipe, expand the propellers, and switch into swimming mode.

5. Spatio-temporal behavioural classification from single-point tracking

In the behavioural-analysis paper, MetaRoundWorm is associated with a method for classifying C. elegans locomotion when only a single tracking point—the worm’s body center—is available. The central motivation is that full-body posture reconstruction becomes unreliable in dense populations where worms overlap or occlude one another, whereas center-point tracking remains feasible. The paper argues that a single-point trajectory still contains enough spatio-temporal structure to recover meaningful behavioural states, and compares a hand-designed “atomic” classification with an unsupervised automatic pipeline (Antonic et al., 30 Sep 2025).

The automatic pipeline operates on 2D center-point trajectories subsampled to 4 Hz. Points with speed above $V_{\text{diag}} = 0.8 \text{ mm/s}$1 are discarded as imaging artifacts. Three feature groups are extracted: kinematic features (speed, angle change, interpolated or resampled angle change with respect to worm body length), spatial features (angular concordance and logarithm of curvature on a sliding window of 10 frames backward and forward), and time-dependent features (lagged speed, lagged angle change, lagged resampled angle change with lag $V_{\text{diag}} = 0.8 \text{ mm/s}$2 frames). Curvature for each crawl segment $V_{\text{diag}} = 0.8 \text{ mm/s}$3 is defined as

$V_{\text{diag}} = 0.8 \text{ mm/s}$4

with a sentinel $V_{\text{diag}} = 0.8 \text{ mm/s}$5 when segment speed is essentially zero. Angular concordance is

$V_{\text{diag}} = 0.8 \text{ mm/s}$6

After feature extraction, directional-change features in $V_{\text{diag}} = 0.8 \text{ mm/s}$7 are converted to absolute values so that only change magnitude is retained. Dimensionality reduction is performed using UMAP with $V_{\text{diag}} = 0.8 \text{ mm/s}$8, $V_{\text{diag}} = 0.8 \text{ mm/s}$9, $^{-1}$0, $^{-1}$1, $^{-1}$2, and $^{-1}$3. Clustering is then performed with KMeans, searching $^{-1}$4 and selecting the optimum by mean silhouette score; this yields $^{-1}$5. The five automatic states are named slow-line (0), straight-turn (1), loop-turn (2), crawl (3), and high-turning (4).

The hand-designed atomic baseline partitions behaviour into crawls and reorientations. Crawls include lines, arcs, and loops; reorientations include sharp turns, pauses, and reversals. Omega turns are omitted because they require full-body outline information. Sharp turns are defined by heading changes of at least $^{-1}$6, pauses by speed below $^{-1}$7, and crawl segments are clustered using KNN with $^{-1}$8 based on logarithm of curvature and angular concordance.

Interpretation of the automatic states uses histogram analysis, an XGBoost classifier, and Shapley values. With stratified group 5-fold cross-validation and Bayesian hyperparameter tuning, the classifier reaches 0.98 overall accuracy, 0.98 macro F1, and class-wise precision, recall, and F1 between 0.98 and 0.99. Shapley analysis indicates that lagged angle-change features are often the most informative variables, supporting the claim that temporal context, not only instantaneous kinematics, structures the classification.

To test behavioural meaning, the paper constructs a probabilistic finite-state agent model. Each state has fitted distributions for speed, angle change, and duration; speed and duration use beta distributions, while angle change uses a mixture of von Mises distributions. State transitions depend on transition frequencies and a weak time-dependent trend:

$^{-1}$9

Model evaluation uses distributional matching of speed and angle-change histograms with the Kolmogorov–Smirnov statistic and mean square displacement over lags up to 1800 s with the average log-likelihood

$^{-1}$0

For both atomic and automatic states, the KS distance between simulated and real distributions remains below 0.078. The MSD average log-likelihood values are -41.14 for atomic, 100 agents; -26.31 for atomic, 1000 agents; -21.64 for automatic, 100 agents; and -22.09 for automatic, 1000 agents. The best overall performance is therefore obtained with the automatic classification, especially at 100 agents.

6. Comparative themes and conceptual boundaries

The cited usages of MetaRoundWorm are unified neither by a common apparatus nor by a common formalism. One usage studies how C. elegans exploits obstacle geometry and contact in a low-Reynolds-number microstructured environment; another uses single-worm microfluidic phenotyping and iterative optimization to discover a four-drug anthelmintic combination; a third concerns a segmented earthworm-inspired robot that combines in-pipe peristaltic crawling with propeller-driven swimming; and a fourth develops an unsupervised spatio-temporal behavioural vocabulary from single-point trajectories [(Majmudar et al., 2010); (Ding et al., 2022); (Fang et al., 2021); (Antonic et al., 30 Sep 2025)].

Even so, some recurring motifs are evident. All four usages are centrally concerned with locomotion or locomotion-derived observables. In the structured-environment study, the decisive variables are hydrodynamic interactions, contact interactions, and geometrical constraints. In the FSC study, behavioural outputs—centroid velocity and track curvature—serve as the objective signal in a model-less control loop. In the robot paper, segmented morphology, peristaltic actuation, and propulsion switching define a dual locomotion architecture. In the behavioural-classification paper, speed, angle change, curvature, angular concordance, and lagged temporal features define the state space. This suggests that the term, as used in this corpus, consistently points toward systems in which worm-like motion is either the phenomenon of interest or the primary measurement channel.

A second boundary concerns modeling assumptions. The bead-chain locomotion model is purely mechanical and excludes sensing or behavior. The FSC framework is explicitly model-less and treats pharmacology and worm biology as a black box. The underwater robot uses a classical kinematics model for crawling and an equivalent dynamic model for swimming. The behavioural-classification pipeline relies on unsupervised clustering and agent-based simulation. A plausible implication is that MetaRoundWorm, in practice, marks a research style that privileges measurable motion, iterative control, or reduced-order representations over detailed organism-level mechanistic closure.

A final conceptual caution follows directly from the corpus: MetaRoundWorm should not be read as naming one settled field-specific artifact. In the available literature, it is a polyvalent label applied to materially different systems and methods. Any technical use of the term therefore requires explicit disambiguation by paper, domain, and operational definition.