MANTA-RAY: Cross-Domain Technical Systems

• MANTA-RAY is a cross-domain research label defining systems like soft manipulation surfaces, astrophysical absorption models, and modular ADAS tracking frameworks.
• The approaches leverage physics-informed design, modular architectures, and control strategies such as PID, PD, and Kalman filtering for efficient operation.
• Empirical results highlight performance metrics including positioning errors below 1 cm, rapid absorption calculations, and high tracking accuracies for marine and automotive applications.

Searching arXiv for the supplied MANTA-RAY-related works to ground the article in current records. MANTA-RAY is a research designation used in several technically distinct ways. In the exact acronymic form, MANTA-RAY denotes “Manipulation with Adaptive Non-rigid Textile Actuation with Reduced Actuation densitY,” a soft, fabric-based manipulation surface with reduced actuator density (Ingle et al., 29 Jan 2026). Closely related spellings also denote “MANTA-Ray,” “Modified Absorption of Non-spherical Tiny Aggregates in the RAYleigh regime,” an analytical model for absorption by non-spherical fractal aggregates (Lodge et al., 2024), and a modular, data-driven multi-object tracking approach built around a Kalman Filter framework in ADAS (Holz et al., 3 Apr 2025). Separately, manta rays are a recurrent biomimetic template for aquatic and aerial robots, and a direct object of marine telemetry and visual re-identification studies (Zhang et al., 2023, Tabata et al., 11 Feb 2026, Nojima-Schmunk et al., 2023, Fernández-Gracia et al., 2024, Moskvyak et al., 2019).

1. Nomenclature and scope

The literature uses the label in both acronymic and biomimetic senses. The exact acronymic forms are not interchangeable: MANTA-RAY in soft manipulation expands to Manipulation with Adaptive Non-rigid Textile Actuation with Reduced Actuation densitY (Ingle et al., 29 Jan 2026); MANTA-Ray in astrophysics expands to Modified Absorption of Non-spherical Tiny Aggregates in the RAYleigh regime (Lodge et al., 2024); and MANTA-RAY in ADAS refers to an overall tracking approach built around SPENT, SANT, and MANTa within a Kalman Filter tracking-by-detection pipeline (Holz et al., 3 Apr 2025). By contrast, several robotic systems are described as manta-ray-inspired rather than acronymic, including a rolled-DEA aquatic robot, a biomimetic underwater robot with servo-driven pectoral fins, and a flapping-wing blimp (Zhang et al., 2023, Tabata et al., 11 Feb 2026, Nojima-Schmunk et al., 2023).

Label	Expansion or inspiration	Domain
MANTA-RAY	Manipulation with Adaptive Non-rigid Textile Actuation with Reduced Actuation densitY	Soft manipulation surface
MANTA-Ray	Modified Absorption of Non-spherical Tiny Aggregates in the RAYleigh regime	Astrophysical dust/haze modeling
MANTA-RAY	Modular KF-integrated tracking approach centered on SPENT, SANT, and MANTa	ADAS multi-object tracking
Manta ray-inspired systems	Median-paired fin locomotion, flapping pectoral fins, or manta-like wing morphology	Aquatic and aerial robotics

This naming pattern makes “MANTA-RAY” a cross-domain research label rather than a single canonical artifact. A plausible implication is that capitalization and expansion are part of the technical identity, not merely stylistic variation.

2. Real manta rays as biological subject and computational target

In marine behavioral ecology, reef manta rays (Mobula alfredi) are studied directly through presence-only acoustic telemetry. A case study at a cleaning station at D’Arros Island, Seychelles used 25 acoustically tagged reef manta rays, detections from one VR2W acoustic receiver, and a roughly 150 m detection radius. The method inferred directed leader–follower relations from lag-time asymmetries using the Kolmogorov–Smirnov arrow, with the final directed network consisting of 12 individuals and 33 directed edges in a single connected component (Fernández-Gracia et al., 2024). The same study reported a clear circadian rhythm with detections most common around noon, fat-tailed interevent times with a power law tail exponent $\alpha \approx 1.38$, and a follower appearance rate higher than the leader’s for roughly the first 200 minutes after leader detection. The inferred structure was not random: females followed males more often than expected, males followed fewer females than expected but with stronger than expected associations, and small individuals following small individuals was much weaker than expected (Fernández-Gracia et al., 2024).

The core dyadic statistic in that study is the signed KS asymmetry,

$$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$

with

$$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$

so that positive or negative sign determines the inferred direction of following (Fernández-Gracia et al., 2024). The method was introduced specifically to avoid arbitrary event windows and the “Gambit of the group.”

Manta rays are also the object of large-scale visual re-identification based on ventral spot patterns. The system developed from Project Manta at the University of Queensland uses a CNN embedding model with triplet loss and online semi-hard triplet mining, rather than closed-set classification (Moskvyak et al., 2019). On the manta-ray dataset, the original collection contained 1730 images of 120 individual manta rays, with 96 individuals for training and 24 individuals for testing. The paper states that, for marine-biological practice, a top-10 accuracy of at least 95% is necessary; the reported system exceeds that threshold, with an InceptionV3 configuration reaching Top-10: 97.78% and AUC: 0.983 in the backbone comparison, while the full manta-ray experiment reports Top-10: 97.03% ± 1.11 and AUC: 0.966 (Moskvyak et al., 2019). A notable technical finding is that, contrary to FaceNet convention, $l_2$-normalizing embeddings hurts performance for this task. The method is explicitly described as generic and not species specific, and was also evaluated on humpback whale flukes (Moskvyak et al., 2019).

3. Biomimetic locomotion inspired by manta rays

Manta-ray locomotion is a recurring engineering template, especially within MPF (Median and/or Pectoral Fin) propulsion. A miniature soft aquatic robot proposed in “Underwater and Surface Aquatic Locomotion of Soft Biomimetic Robot Based on Bending Rolled Dielectric Elastomer Actuators” uses one bending rolled dielectric elastomer actuator (DEA) per fin, converting the in-plane expansion of a rolled tube into out-of-plane bending by bonding one side to a 0.2 mm PDMS constraining film (Zhang et al., 2023). Each actuator is made from a nine-layer dielectric elastomer multilayer sheet of Silikon Addition Farblos 5, with each layer spin-coated to about 31 μm, then rolled into a tube of roughly 4 mm diameter and 25 mm length. The actuator reaches a maximum free-end displacement of about 17 mm, a resonance-like bending-angle peak of $52 \pm 2^\circ$ around 17 Hz, and a blocked force of about 55 mN at 1200 Vpp and 17 Hz (Zhang et al., 2023). Under open-loop locomotion, the robot swims at 57 mm/s or 1.25 body length per second (BL/s) underwater, skates at 64 mm/s or 1.36 BL/s on the water surface, and ascends vertically at 38 mm/s or 0.82 BL/s with a 5 mm × 4 mm × 4 mm float, all at 1300 V and 17 Hz (Zhang et al., 2023). The paper interprets the hydrodynamics in terms of traveling waves along the flexible fins and vortex shedding at the trailing edges.

A larger biomimetic manta-ray robot directed toward underwater autonomy uses pectoral fins actuated by two Futaba RS303MR servo motors per fin for flapping about the X-axis and feathering about the Y-axis (Tabata et al., 11 Feb 2026). The platform measures 360 mm in length, 750 mm in width, 70 mm in height, and 2150 g in weight; the onboard system includes a Raspberry Pi 3B, an MPU9050 IMU, an Arduino Nano, and an LPS33HW pressure sensor (Tabata et al., 11 Feb 2026). The flapping and feathering trajectories are given by

$$\theta_{fl} = \theta_{\rm{flmax}} \sin(2\pi tf), \qquad \theta_{fe} = \theta_{\rm{femax}} \sin\left(2\pi tf - \frac{\pi}{2}\right),$$

with a 90° phase difference explicitly chosen to maximize thrust (Tabata et al., 11 Feb 2026). In surface swimming with 30° flapping, 45° feathering, and 0.75 Hz, the robot stabilizes at about 20 cm/s and traverses 500 cm in about 20 seconds under PD control; in a diving-motion experiment with a target depth of 10 cm, the average swimming speed during the straight segment is 22 cm/s (Tabata et al., 11 Feb 2026). The paper also documents that PD control alone is insufficient against large disturbances such as bottom collision, and that pitch motion causes IMU integration error during diving (Tabata et al., 11 Feb 2026).

Manta-ray inspiration has also been adapted to lighter-than-air robotics in “Manta Ray Inspired Flapping-Wing Blimp” (Nojima-Schmunk et al., 2023). The vehicle, called Flappy, is built from two 91.4 cm diameter ellipsoidal balloons filled with helium, together providing 136 g of lift, and carries two flapping wings, one tail, three 9 g servos, an ESP32 Feather, and a 2S 300 mAh LiPo battery (Nojima-Schmunk et al., 2023). The best-performing wing in the parametric thrust study was Stiff, with $AR = 1.0$, $\gamma = 0.25$, and a concave trailing edge; at 90° amplitude and 1.25 Hz, it produced 17.3 g average thrust (Nojima-Schmunk et al., 2023). With that wing, the blimp reached 1.1 m/s maximum speed and 2420 m maximum range. The paper reports a 68% increase in range over an otherwise identical propeller-based platform (Nojima-Schmunk et al., 2023).

4. MANTA-RAY as a modular soft manipulation surface

The exact acronymic MANTA-RAY in robotics names a soft, fabric-based surface with reduced actuator density designed for manipulation without direct grasping (Ingle et al., 29 Jan 2026). Earlier work used a single module supported by four actuators; the multi-modular extension presents a distributed architecture in which local modules share boundary actuators and objects are transferred by object passing between modules (Ingle et al., 29 Jan 2026). The physical platform is about 1 m × 1 m, with resting height about 0.7 m, 9 actuators (A0–A8) arranged in a 3×3 grid, and 4 modules (M0–M3) in a 2×2 tiling. Actuator spacing is 0.5 m, and each actuator has about 0.4 m vertical stroke. The actuators use Nema 23 stepper motors, a pulley-belt mechanism, linear guides, Arduino Uno, CNC V3 shield, A4988 stepper driver, and an AS500 magnetic encoder with 12-bit resolution. The fabric is 100% polyester, approximately 1.2 × 1.2 m, with each module using a 0.6 × 0.6 m hanging section; motion tracking uses OptiTrack at up to 200 Hz (Ingle et al., 29 Jan 2026).

A central claim of the platform is that useful manipulation can be achieved with much lower actuator density than dense arrays. The paper states that the system can achieve an object-to-actuator size ratio as low as 0.01, meaning objects can be about 100× smaller than actuator spacing (Ingle et al., 29 Jan 2026). Object passing is implemented by raising the actuators belonging exclusively to the current module while lowering the shared boundary actuators and the neighboring module, creating a local slope toward the target module.

The low-level controller is a geometric transformation-driven PID controller. Positional error is converted into desired surface tilt, the tilt defines a plane, and the plane is mapped directly to actuator heights: $$\mathbf{T}_i = (\sin \theta_{zx}, \sin \theta_{zy}, \cos \theta_{zx}),$$

$$\mathbf{T}_i \cdot (\mathbf{r} - \mathbf{r}_0) = 0, \qquad \mathbf{r}_0 = (0,0,z_0),$$

$$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$0

$$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$1

The additive term $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$2 introduces small vibrations to overcome static friction (Ingle et al., 29 Jan 2026).

The platform was evaluated in simulation on 2×2 and 3×3 module configurations and on a physical 2×2 prototype. Tested objects included a sphere, cube, disk, apple, cylinder, egg, and dice (Ingle et al., 29 Jan 2026). In passing experiments from M0 to M2 and back, repeated three times, the reported mean standard deviations included Sphere: $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$3, $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$4, $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$5 m and Dice: $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$6, $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$7, $$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$8 m; across all objects, mean positional deviation stayed within about 0.03 m (Ingle et al., 29 Jan 2026). In target-reaching tests, all objects reached their targets; the experimental section reports mean positioning error less than 0.02 m, while the conclusion states below 1 cm average positioning error over the tested area (Ingle et al., 29 Jan 2026). In hardware multi-object manipulation, a sphere and a disk were manipulated simultaneously at an average control frequency of 20 Hz, with both staying within the 3 cm threshold (Ingle et al., 29 Jan 2026).

5. MANTA-Ray and related systems in modeling and tracking

In astrophysics, MANTA-Ray is a fast analytical model for the absorption efficiency of non-spherical fractal aggregates in the long-wavelength or Rayleigh limit (Lodge et al., 2024). Its central relation is

$$A_{\mathrm{KS}(P_1, P_2)} = P^{(1)}(\tau) - P^{(2)}(\tau),$$9

where $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$0 is a multiplicative enhancement factor depending on refractive index and fractal dimension (Lodge et al., 2024). The model is validated for

$$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$1

with homogeneous composition and

$$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$2

The paper states that MANTA-Ray calculates absorption efficiencies within 10–20% of DDA while being $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$3 times faster, and emphasizes that treating non-spherical aggregates as spheres can be catastrophically inaccurate: for $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$4, the spherical approximation can underestimate absorption by a factor of 1,000, with reported errors up to 31,000–110,000% maximum across the tested range (Lodge et al., 2024).

In underwater computer vision, MANTA is a physics-informed framework for underwater single-object tracking that combines dual-positive contrastive learning, Beer–Lambert-law-based physics augmentations, and a multi-stage tracking pipeline (Srinath et al., 28 Nov 2025). The representation is trained with temporal positives and Beer–Lambert augmented positives, while the deployed system combines RF-DETR, OC-SORT, and a secondary association stage based on geometric consistency and appearance similarity (Srinath et al., 28 Nov 2025). The paper also introduces Center-Scale Consistency (CSC) and Geometric Alignment Score (GAS) as geometry-aware metrics. On UOT32, the reported scores include Success AUC: 0.7306, [email protected]: 0.8848, +11.5% [email protected], and +4.2% mGAS over comparison methods; on runtime, MANTA: 37 FPS with 45.4M parameters (Srinath et al., 28 Nov 2025).

In ADAS, MANTA-RAY is a modular, data-driven multi-object tracking system that preserves a classical Kalman Filter tracking-by-detection architecture while replacing selected subroutines with compact neural networks (Holz et al., 3 Apr 2025). The three named modules are SPENT for trajectory prediction, SANT for single-association, and MANTa for multi-association (Holz et al., 3 Apr 2025). Each network contains less than 50k trainable parameters. On KITTI, SPENT reports Testing RMSE: 0.029 versus 0.066 for a standard KF; SANT reaches 95% assignment accuracy on a test set of 391 samples; MANTa reaches 95% accuracy for 1 to 6 tracks per timestamp but only 14% accuracy for 7 to 16 tracks, with 80% overall average (Holz et al., 3 Apr 2025). The paper explicitly states that the system is not a single monolithic tracker, but a learned modular augmentation of a conventional KF-based tracker.

The broader MANTA naming stem also appears in other technical systems. In neutron instrumentation, MANTA means Multi-Analyzer Neutron Triple-Axis, a planned cold-neutron TAS for HFIR using a multiplexed prismatic analyzer concept with $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$5 angular channels, $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$6 detectors per channel, $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$7 analyzer stations per channel, and an idealized simultaneous count of 832 points in $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$8 (Desai et al., 2023). In metropolitan traffic simulation, MANTA means Microsimulation Analysis for Network Traffic Assignment, a GPU-parallel platform that simulates the nine-county San Francisco Bay Area morning period in 4.6 minutes for the microsimulation component, using 0.5-second timesteps (Yedavalli et al., 2020). These uses are terminologically related but technically separate.

6. Recurring themes, limitations, and misconceptions

A common misconception is that MANTA-RAY denotes a single biomimetic robot. Current literature does not support that reading. The exact label refers at minimum to a soft manipulation surface (Ingle et al., 29 Jan 2026), an astrophysical absorption model (Lodge et al., 2024), and a Kalman-integrated tracking framework (Holz et al., 3 Apr 2025), while manta-ray inspiration separately motivates underwater and aerial robots (Zhang et al., 2023, Tabata et al., 11 Feb 2026, Nojima-Schmunk et al., 2023). Another common misconception is that these systems are uniformly end-to-end learned. Several are explicitly not: the soft manipulation platform uses a geometric transformation-driven PID controller (Ingle et al., 29 Jan 2026), the ADAS tracker preserves modular KF structure (Holz et al., 3 Apr 2025), and the aquatic robots rely on open-loop or PD control rather than learned policies (Zhang et al., 2023, Tabata et al., 11 Feb 2026).

The limitations are strongly domain specific. The rolled-DEA aquatic robot remains tethered by a high-voltage supply, exhibits wiring-related imbalance, and requires more detailed hydrodynamic modeling, resonance analysis, attitude control, and eventual autonomous surface-water transition (Zhang et al., 2023). The underwater autonomous biomimetic robot is vulnerable to bottom collision, and its diving accuracy degrades because pitch motion causes cumulative IMU integration error (Tabata et al., 11 Feb 2026). The soft manipulation surface shows motion variability that depends strongly on object geometry, and parallel manipulation can suffer from interference when adjacent modules deform simultaneously (Ingle et al., 29 Jan 2026). The astrophysical MANTA-Ray model is only validated in the long-wavelength limit, assumes homogeneous composition, and is not intended for $$\tau = \arg\max_t \left| P^{(1)}(t) - P^{(2)}(t) \right|,$$9 in the main model (Lodge et al., 2024). The underwater tracker depends on UDepth and uses fixed thresholds in secondary association (Srinath et al., 28 Nov 2025). The ADAS MANTa module is strongly affected by data imbalance, with most training samples containing only one to six tracks (Holz et al., 3 Apr 2025). The wildlife re-identification system still requires a manual bounding box around the pattern of interest, and performance drops for sparse markings, heavy occlusion, or uninformative views (Moskvyak et al., 2019).

Taken together, these works suggest a recognizable research pattern rather than a single technology. The recurring themes are modularity, physics-aware design, and functional efficiency under constrained actuation or computation: reduced actuator density in textile manipulation (Ingle et al., 29 Jan 2026), a multiplicative correction to Rayleigh absorption rather than full DDA (Lodge et al., 2024), compact subnetworks embedded in a KF tracker (Holz et al., 3 Apr 2025), and compact flapping or DEA-based propulsion mechanisms that preserve manta-like fin kinematics (Zhang et al., 2023, Tabata et al., 11 Feb 2026, Nojima-Schmunk et al., 2023). The term “MANTA-RAY” is therefore best understood as a family of domain-specific technical constructs linked by naming, and only sometimes by direct biological inspiration.