Animal-in-Motion (AiM) Research

• Animal-in-Motion (AiM) is a research orientation that models animal movement as a structured dynamical phenomenon, emphasizing kinematics, interactions, and latent behavioral states.
• AiM employs diverse methodologies including maximum-entropy stochastic models, geometric manifold learning, and probabilistic hybrid frameworks to capture continuous-time, spatial, and state-dependent dynamics.
• AiM benchmarks and systems—from 4D quadruped reconstruction to autonomous airship tracking—demonstrate its practical applications in advancing motion inference, control, and behavior analytics.

Searching arXiv for the supplied AiM-related papers and benchmark context. Animal-in-Motion (AiM) denotes a cluster of research programs centered on representing, measuring, reconstructing, generating, and interpreting animal movement as a structured dynamical phenomenon rather than as trajectory geometry alone. In the literature considered here, the term is used in several related senses: a maximum-entropy theory of continuous-time movement (Fleming et al., 2014), an information-theoretic framework for dynamical coupling in moving groups (Richardson et al., 2013), a geometric learning formulation for estimating motion submanifolds (Powell et al., 2020), a probabilistic hybrid model of behavioral-state switching and locomotion (Bodova et al., 2017), a web-collected benchmark for markerless 4D quadruped reconstruction (Zhao et al., 3 Nov 2025), and an autonomous airship-based motion-capture system for wild horses (Price et al., 2024). This plurality suggests that AiM is best understood as a research orientation in which animal motion is treated simultaneously as kinematics, interaction, embodiment, and latent structure.

1. Terminological scope and recurring formulations

Across recent work, AiM does not refer to a single canonical dataset or method. Instead, it names several technically distinct but conceptually adjacent efforts that all treat motion as the primary substrate for inference, control, or reconstruction. In one strand, AiM is a theory of continuous-time stochastic movement obtained by maximum entropy under finite kinematic autocorrelation constraints (Fleming et al., 2014). In another, it is a framework for inferring directed influence between animals from trajectory data by conditional mutual information rather than spatial order (Richardson et al., 2013). In computer vision, AiM is also the name of a benchmark of 230 manually curated sequences with 11,061 frames for 4D quadruped reconstruction, derived from a larger web-collected corpus of 29,979 animal video clips totaling 2,046,414 frames (Zhao et al., 3 Nov 2025). In robotics and field systems, AiM appears as an autonomous aerial observation platform built from cooperative airships for tracking and visually recording wild horses (Price et al., 2024).

AiM usage	Representative formulation	Representative source
Continuous-time movement theory	Maximum-entropy states for animal movement	(Fleming et al., 2014)
Interaction analysis	Dynamical coupling and directional information flow	(Richardson et al., 2013)
Geometric learning	Estimation of animal motion submanifolds	(Powell et al., 2020)
Hybrid behavior-motion modeling	PDMP with GLM switching rates	(Bodova et al., 2017)
4D reconstruction benchmark	Web-collected in-the-wild quadruped benchmark	(Zhao et al., 3 Nov 2025)
Autonomous capture system	Airship formations for motion capture and behavior analysis	(Price et al., 2024)

A recurring theme across these usages is that motion is not treated as a by-product of static labels. Instead, movement becomes the object from which one infers leadership, behavioral state, morphology-conditioned dynamics, scene interaction, or internal state. A plausible implication is that AiM functions less as a narrow subfield label than as a common problem formulation linking movement ecology, statistical learning, computer vision, robotics, and embodied AI.

2. Information flow, coupling, and collective motion

A central AiM problem is the inference of who influences whom in a moving group when relative position is misleading. The meerkat study of dynamical coupling addresses this problem using high spatio-temporal resolution GPS trajectories sampled at 1 Hz and an information-theoretic framework based on mutual information and conditional mutual information (Richardson et al., 2013). The symmetric quantity

$I(A_t;B_{t+lag})$

measures predictive dependence, while directional influence is captured by

$I(A_t;B_{t+lag}\mid B_t),$

denoted compactly as $I(A \rightarrow B)$. To summarize asymmetry, the paper defines

$$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$

Positive values indicate that $A$ influences $B$; negative values indicate the reverse.

The methodological logic is explicit. For each pair of animals, trajectories are converted into time series of displacements; surrogate trajectories preserve start and end points and within-trajectory structure while removing coupling; directional information flow is computed in both directions; and randomization tests determine whether a coupling event is statistically significant. The use of sliding windows makes the analysis time-resolved, revealing intermittent fluctuation of the coupling strength and alternation in the coupling direction within foraging bouts (Richardson et al., 2013).

This framework directly challenges a common simplification in collective-motion analysis: the equation of front position with leadership. Relative longitudinal position is computed by inferring a smoothed group trajectory from the pair’s barycenter and projecting each animal onto it, but the resulting comparison shows that coupling direction is not a simple monotonic function of position. Coupling can occur from frontal to trailing individuals and vice versa, with peaks both when the driver is at the front and when the driver is at the rear. Around the center zone there is often little net directionality. The paper therefore rejects the claim that front position alone determines who leads (Richardson et al., 2013).

The broader AiM significance is methodological. By focusing on information flow rather than correlation, the framework is designed to detect nonlinear or lagged dependencies that may not manifest as simple alignment correlations. The paper explicitly argues that correlation-based methods can under-detect coupling, creating a risk of false negatives. This suggests an AiM perspective in which collective motion is modeled as a network of transient directed influences rather than a static ordering of individuals.

3. Continuous-time, geometric, and probabilistic models of motion

A second major AiM strand formalizes animal movement as a mathematical object with explicit regularity, continuity, and learning guarantees. In the maximum-entropy formulation, the trajectory is treated as a stochastic process, and only the mean and a finite number of kinematic autocorrelation constraints are imposed. The central spectral-density class is

$$\tilde{\boldsymbol{\sigma}}(f) = \left[ \sum_{k=0}^{K} (2\pi f)^{2k}\,\boldsymbol{\lambda}_k \right]^{-1},$$

which naturally includes Brownian motion, Ornstein-Uhlenbeck motion, integrated Ornstein-Uhlenbeck motion, the hybrid OUF model, and a central-place foraging model (Fleming et al., 2014). The theory further states that Langevin equations must obey a fluctuation-dissipation theorem to generate processes in this maximum-entropy class. In that sense, AiM here is a unifying theory of continuous-time movement rather than a collection of disconnected stochastic models.

A complementary geometric formulation treats a motion regime as lying on an unknown configuration submanifold $Q\subset X\approx \mathbb{R}^d$, parameterized by a known smooth compact connected Riemannian manifold $S$ through an unknown map $\gamma:S\rightarrow Q\subset X$ (Powell et al., 2020). Data are i.i.d. samples $I(A_t;B_{t+lag}\mid B_t),$0, the ideal risk is

$I(A_t;B_{t+lag}\mid B_t),$1

and the empirical risk minimizer over a finite-dimensional approximation space is

$I(A_t;B_{t+lag}\mid B_t),$2

Under the paper’s approximation-class assumptions, each component satisfies the distribution-free bound

$I(A_t;B_{t+lag}\mid B_t),$3

Experimentally, the method is illustrated on reptile motion data, including marked lizards and Anolis sagrei, with three high-speed Photron FASTCAM cameras at 500 fps (Powell et al., 2020).

Where the previous two formulations emphasize continuity or geometry, the probabilistic hybrid framework models individual and collective behavior as a piecewise-deterministic Markov process. Each animal occupies a behavioral state $I(A_t;B_{t+lag}\mid B_t),$4, follows a state-specific deterministic motion law

$I(A_t;B_{t+lag}\mid B_t),$5

and switches stochastically between states with rates expressed as a Generalized Linear Model (Bodova et al., 2017). With exponential link $I(A_t;B_{t+lag}\mid B_t),$6, the transition rate includes single-animal covariates and pairwise interaction covariates. Forward simulation uses a variant of Gillespie’s Stochastic Simulation Algorithm; inference is maximum-likelihood and tractably solvable by gradient descent; and model selection is used to identify factors that modulate behavioral-state switching (Bodova et al., 2017).

Taken together, these three formulations exemplify a broad AiM mathematical agenda. Maximum entropy supplies least-assumptive continuous-time priors; manifold learning supplies coordinate-free low-dimensional representations with convergence rates; and PDMP-GLM models supply stochastic switching between discrete behavioral modes with state-dependent motion laws. This suggests that AiM is not committed to a single ontology of motion, but rather to the principle that motion should be modeled at the level appropriate to its continuity, geometry, and latent behavioral structure.

4. Data infrastructure and 4D reconstruction benchmarks

In computer vision, AiM is also a concrete benchmark and data-engineering effort for markerless 4D quadruped reconstruction from in-the-wild video (Zhao et al., 3 Nov 2025). The underlying pipeline begins from 23 common animal categories and automatically mines YouTube videos using GPT-generated search terms, Selenium WebDriver, and pytubefix. Shot detection is performed with PySceneDetect, clips shorter than 30 frames are discarded, CLIP and CLIPScore filter clips against prompts such as “A photo of {category},” and all clips are downsampled to 10 fps. Grounded-SAM-2, combining GroundingDINO and SAM-2, provides object detection and tracking. The resulting tracks are filtered for overlapping instances, low resolution, truncation, inconsistent tracking, and other failures before forming square object-centric crops (Zhao et al., 3 Nov 2025).

The processed resource contains 29,979 animal video clips totaling 2,046,414 frames. From these, Animal-in-Motion is constructed as an evaluation-only benchmark of 230 manually curated sequences with 11,061 frames, using 10 videos per category. Sequences are retained only if they exhibit minimal heavy occlusion by other objects, recognizable and smooth body motion, smooth camera motion, accurate masks without missing body parts, and keypoints that are accurate and temporally stable (Zhao et al., 3 Nov 2025).

The benchmark includes auxiliary annotations and derived features: instance masks, 2D keypoints from ViTPose++, DINOv2 features, SEA-RAFT optical flow, Depth Anything V2 depth estimates, and occlusion boundary estimates derived from depth at dilated versus eroded mask boundaries (Zhao et al., 3 Nov 2025). The task is to estimate a sequence of 3D poses and shapes from a video

$I(A_t;B_{t+lag}\mid B_t),$7

with auxiliary inputs

$I(A_t;B_{t+lag}\mid B_t),$8

via a method

$I(A_t;B_{t+lag}\mid B_t),$9

that outputs posed 3D shapes $I(A \rightarrow B)$0 (Zhao et al., 3 Nov 2025).

A central finding is evaluative rather than merely algorithmic. On AiM, SMALify achieves the best 2D-based metrics, including IoU $I(A \rightarrow B)$1, [email protected] $I(A \rightarrow B)$2, [email protected] $I(A \rightarrow B)$3, [email protected] $I(A \rightarrow B)$4, [email protected] $I(A \rightarrow B)$5, and MPJVE $I(A \rightarrow B)$6, while model-free methods such as 3D-Fauna and the sequence-optimized 4D-Fauna often yield more natural reconstructions but lower scores (Zhao et al., 3 Nov 2025). The paper argues that 2D metrics favor model-based methods even when their 3D shapes are unrealistic, revealing a gap between current evaluation practice and actual 3D plausibility.

The 4DEquine work uses AiM precisely in this benchmarking role and reports state-of-the-art performance on real-world horse and zebra subsets while disentangling motion reconstruction from appearance reconstruction (Lyu et al., 10 Mar 2026). On the AiM horse motion subset, AniMoFormer achieves [email protected] $I(A \rightarrow B)$7, [email protected] $I(A \rightarrow B)$8, and Accel $I(A \rightarrow B)$9. For appearance reconstruction on the AiM horse subset, 4DEquine reports PSNR $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$0, SSIM $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$1, and LPIPS $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$2; on the zebra subset, it reports PSNR $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$3, SSIM $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$4, and LPIPS $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$5 (Lyu et al., 10 Mar 2026). The paper attributes these gains to temporal modeling plus post-optimization for motion and to a feed-forward Gaussian avatar reconstruction network for appearance.

The benchmark literature therefore introduces a second common misconception. High silhouette or keypoint agreement is not equivalent to high-fidelity 3D reconstruction. AiM is used here not only as data, but as a diagnostic instrument exposing the difference between 2D fit and 3D realism.

5. Capture systems, embodied environments, and synthetic testbeds

AiM also names systems in which motion is acquired or studied through autonomous platforms and embodied agents. The airship-based AiM system is designed for tracking, following, and visually recording wild horses from multiple angles, using cooperative formations rather than a single UAV (Price et al., 2024). The prototype is a 5.5 m long, 3.6 m³ single-hull airship with about 430 g envelope mass, about 1800 g payload capacity, two 4S 5500 mAh LiPo batteries totaling 125 Wh, two fixed brushless propellers in pull configuration, and a Jetson TX2 onboard computer on an AUVIDEA J-120 board (Price et al., 2024). Its sensing stack includes two Logitech Brio 4K webcams, a flight controller based on an OpenPilot Revolution STM32 board with IMU, magnetometer, and barometric pressure sensor, GPS, an airspeed sensor, Wi-Fi, and a 1 TB SSD.

The control architecture is layered. A ROS-enabled version of Librepilot provides low-level stabilization through an EKF and a cascaded PID controller for controlled airspeed, climb/sink rate, heading/turn-rate control, position hold, and waypoint navigation. A high-level model predictive controller estimates wind and generates subject-centered trajectories that keep the subject centered in the camera frame while maintaining formation geometry, collision-avoidance constraints, and a prohibition against direct overflight of the subject (Price et al., 2024). The simulation environment is ROS + Gazebo with aerodynamic physics, wind and turbulence disturbance models, sensor noise, actuator effects, and hardware-in-the-loop capability.

Field experiments in August 2023 at Hortobágy National Park established practical operating constraints. Ground handling was identified as the hardest part of deployment; the vehicle was trimmed about 300 g heavier than air for safety; practical wind conditions were about 6–8 m/s with gusts over 10 m/s on some days; a minimum airspeed of 4 m/s was necessary to maintain sufficient vertical control authority; practical Wi-Fi range was about 150 m; and the system achieved autonomous tracking of wild horses after only four days from maiden flight to field autonomy (Price et al., 2024). The paper reports that wildlife observers judged the acoustic footprint to be much lower than that of a comparable drone and that the horses were not visibly disturbed.

A different embodied AiM formulation appears in the Articulated Animal AI Environment for Animal Cognition, which extends the original AnimalAI benchmark from a simple spherical agent to a soft-body, multi-jointed animal-like creature (Lucas et al., 2024). The articulated agent has four thighs, four legs, and eight joints total; each joint can rotate in the $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$6 direction from $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$7 to $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$8 degrees and in the $$D(A \rightarrow B)=\frac{I(A \rightarrow B)-I(B \rightarrow A)}{I(A \rightarrow B)+I(B \rightarrow A)}.$$9 direction from $A$0 to $A$1 degrees; the head has mass $A$2; and each thigh and leg has mass $A$3 (Lucas et al., 2024). The environment adds integrated curriculum training and evaluation tools, randomized tests, joint-rotation or joint-velocity action modes, camera or raycast observations, and task suites extending from L0 Initial Food Contact to L11 Body Awareness. The curriculum samples task cells using

$A$4

with rolling moving average and variance over the last 10 episodes (Lucas et al., 2024).

These system papers place AiM at the intersection of sensing, control, and embodiment. One branch studies real animals through long-endurance subject-centered observation; the other studies animal-like cognition in a limbed agent whose movement is integral to the task structure. In both cases, locomotion is not incidental to cognition or data collection; it is the mechanism through which the system functions.

6. Generation, transfer, virtual animals, and welfare-oriented interpretation

Recent AiM-adjacent work expands from analysis and reconstruction to controllable synthesis, motion transfer, and virtual embodiment. In robotics, an end-to-end framework learns steerable imitation controllers from unstructured real animal motion data by combining offline kino-dynamic motion retargeting, a hyperspherical VAE with a mixture-of-experts decoder, an RL latent-space policy, and a residual RL tracking controller (Kang et al., 1 Jul 2025). The VAE is trained on 49-dimensional state transitions, uses an 18-dimensional latent space with a von Mises-Fisher prior, and is optimized with

$A$5

with $A$6. A command-conditioned synthesis policy receives $A$7 and produces latent actions projected onto the hypersphere, enabling smooth transitions such as Gallop to Trot to Pace as commanded forward speed decreases from about $A$8 m/s to $A$9 m/s to $B$0 m/s (Kang et al., 1 Jul 2025).

For animation and motion transfer, one line of work emphasizes species-specific habitual behavior rather than mere skeletal alignment. The habit-preserved cross-category motion-transfer framework uses a VQ-VAE, a category-specific habit encoder, and an LLM-based text encoder, and is evaluated on DeformingThings4D-skl with 21 animal categories, 787 4D motion sequences, and 29,505 frames (Zhang et al., 10 Jul 2025). Its reported best variant yields FID $B$1, Intra-FID $B$2, downstream-task score $B$3, Diversity $B$4, 1-NNA $B$5, and MPJPE $B$6 (Zhang et al., 10 Jul 2025). A complementary topology-agnostic text-to-motion framework introduces OmniZoo with 140 species, 32,979 sequences, 2,488,539 frames, and 46K text descriptions, and uses a topology-aware skeleton embedding module together with a generalized autoregressive token model to generate motions for arbitrary skeletal topologies (Chen et al., 11 Dec 2025). Its main table reports FID $B$7, Diversity $B$8, MatchingScore $B$9, MultiModality $$\tilde{\boldsymbol{\sigma}}(f) = \left[ \sum_{k=0}^{K} (2\pi f)^{2k}\,\boldsymbol{\lambda}_k \right]^{-1},$$0, and R-Precision values $$\tilde{\boldsymbol{\sigma}}(f) = \left[ \sum_{k=0}^{K} (2\pi f)^{2k}\,\boldsymbol{\lambda}_k \right]^{-1},$$1, $$\tilde{\boldsymbol{\sigma}}(f) = \left[ \sum_{k=0}^{K} (2\pi f)^{2k}\,\boldsymbol{\lambda}_k \right]^{-1},$$2, and $$\tilde{\boldsymbol{\sigma}}(f) = \left[ \sum_{k=0}^{K} (2\pi f)^{2k}\,\boldsymbol{\lambda}_k \right]^{-1},$$3 for R@1, R@2, and R@3 (Chen et al., 11 Dec 2025).

Virtual-animal systems push AiM toward immersive rendering and scene-aware animation. ARTEMIS combines articulated neural appearance synthesis with motion synthesis for furry animals, using a dynamic octree-based neural volume, voxel-level deformation by explicit skeletal warping, and a motion-control module based on captured real animal skeletal motion (Luo et al., 2022). Virtual Pets instead reconstructs foreground deformable NeRFs and background static NeRFs from monocular internet videos, then trains a conditional 3D motion model with separate VAEs for trajectory and articulation to generate cats moving plausibly in 3D rooms (Cheng et al., 2023). These works suggest an AiM trajectory in which motion, appearance, articulation, and environment become jointly modeled rather than separately layered.

At the interpretive end of the spectrum, the survey of computer-vision-based recognition of animal pain and affective states argues that one must go “deeper than tracking” (Broomé et al., 2022). Tracking and pose tools such as DeepLabCut, LEAP, DeepPoseKit, and idtracker.ai provide the substrate, but the paper stresses that being able to track the intricate movements of an animal does not mean its behavior is understood. The survey distinguishes single-frame, frame-aggregation, and spatiotemporal video modeling; emphasizes that temporal dynamics matter for questions such as whether a blink differs from a half-blink; and identifies pain, stress, anticipation, frustration, and positive affect as downstream targets of inference (Broomé et al., 2022). This introduces a third recurring misconception: tracking is necessary, but not sufficient, for behavior understanding.

Taken together, these generative and interpretive strands show how AiM has expanded beyond measurement. Motion is now used to drive quadruped robots, transfer species-specific habits across categories, generate topology-aware animal actions from text, animate neural pets in real time, and support welfare-oriented inference from facial and bodily behavior. A plausible implication is that AiM increasingly functions as an interface layer between raw movement and higher-level constructs such as identity, habit, affect, and controllable embodiment.