Squirrels in Ecology, Robotics & Cryptography

• Squirrels are small mammals studied as multifaceted systems in ecology, robotics, and even cryptography.
• Research integrates occupancy modeling, spatial competition, and nonlinear physiological analysis to reveal behavioral adaptations and ecological dynamics.
• Their biological and engineered traits inform the design of agile drones, legged robots, and post-quantum signature schemes.

Squirrels, in the contemporary technical literature, are studied less as a single descriptive natural-history category than as a set of empirically rich systems for statistical ecology, nonlinear physiology, demography, locomotion, robotics, infrastructure reliability, and even cryptographic nomenclature. The research surveyed here centers on European red squirrels (Sciurus vulgaris), eastern gray squirrels (Sciurus carolinensis), fox squirrels (Sciurus niger), Arctic ground squirrels, Columbian ground squirrels (Urocitellus columbianus), and flying squirrels, and it treats squirrels both as biological subjects and as comparative templates for hierarchical binary regression, partially observed dynamical inference, spatial spread models, observer-aware control, and morphing-wing robotic design (Scharf et al., 2019, FitzGerald et al., 1 Apr 2025, Armesto et al., 3 Apr 2026, Banegas et al., 3 Sep 2025).

1. Species, settings, and observational regimes

The squirrel literature represented here is organized around distinct empirical settings rather than a single canonical observational protocol. European red squirrels are analyzed through occupancy data collected at 265 Swiss sites with repeated visits and landscape and detection covariates; Scottish red and grey squirrels are tracked on a 10,057-cell spatial grid using public sighting records; Arctic ground squirrels are studied through body-temperature time series recorded from free-ranging individuals during hibernation; Columbian ground squirrels are followed across 32 years for phenology, body mass, breeding success, and survival; and fox, eastern gray, and red squirrels appear in field experiments on leaping, caching, recovery, and audience-sensitive behavior (Scharf et al., 2019, Gramain, 2023, FitzGerald et al., 1 Apr 2025, Tamian et al., 7 Jul 2025, Armesto et al., 3 Apr 2026).

These settings isolate different facets of squirrel biology. Some studies emphasize hidden-state inference under imperfect detection, as in Swiss occupancy analyses. Others emphasize long-horizon ecological dynamics, as in the Scottish spread model or the Columbian ground squirrel demographic series. Still others treat squirrel behavior as a high-bandwidth control-and-memory problem: arboreal locomotion, scatter-hoarding, delayed retrieval, and strategic caching in the presence of observers. A separate engineering line treats flying squirrels and tree-jumping squirrels as templates for agile drones and perch-capable robots, indicating that squirrel research now spans both biological and synthetic systems (Xu et al., 2024, Kang et al., 13 Apr 2025).

2. Occupancy, competition, and spatial spread

A central statistical treatment of squirrels appears in hierarchical binary regression for European red squirrel occupancy in Switzerland. In that framework, binary detections $y_{ij}$ at site $i$ and visit $j$ are generated through latent occupancy-abundance and detection variables,

$y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$

The latent abundance variable $z_i$ can follow a Poisson law,

$$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$

while the detection process is modeled by

$$u_{ij}\mid \boldsymbol{\alpha}\sim \mathcal{N}(\mathbf{w}_{ij}'\boldsymbol{\alpha},1),$$

so that $r_{ij}=\Phi(\mathbf{w}_{ij}'\boldsymbol{\alpha})$ and $\psi_i=\Pr(z_i>0)=1-\exp(-\lambda(\mathbf{x}_i))$. The paper also develops a Poisson model with a spatially structured random effect $\zeta_i$ with $i$0, where $i$1 is a CAR precision matrix. Across naive, Kéry-Royle logistic/probit, and Poisson models with linear or quadratic covariate expansions, predictive performance under cross-validation was similar across all models except the naive one, whereas the Poisson-based constructions yielded additional ecological interpretation through abundance mapping and log-linear effects on abundance (Scharf et al., 2019).

At the population-dynamical scale, squirrel competition has been formalized through two-population and three-population systems with explicit safety niches. In the two-population case, the red native population $i$2 and grey exotic population $i$3 satisfy

$i$4

where $i$5 is the fraction of red squirrels in a safety niche. In the three-population extension, an additional red indigenous population $i$6 is introduced with a second niche fraction $i$7. The key analytical object is the separatrix manifold, a curve in 2D or a surface in 3D separating basins of attraction of stable equilibria. The paper reconstructs that manifold by detecting points through trajectory comparison and bisection, refining them by local averaging, and interpolating them with a Partition of Unity method using Wendland’s compactly supported $i$8 radial basis function

$i$9

Within this formulation, increasing $j$0 or $j$1 shifts the separatrix in favor of protected red populations by reducing effective exposure to interspecific competition (Cavoretto et al., 2014).

A still larger-scale representation appears in the discrete spatio-temporal model for Scottish red and grey squirrels. Scotland is discretized into 10,057 cells of about $j$2, with annual updates composed of a local Lotka–Volterra competition step and a dispersal step to adjacent and diagonal neighbors. Cell-specific attractiveness depends on altitude and forest type, with total attractiveness given by the product of altitude and woodland attractiveness. The model is initialized from public sightings and validated against records from 2006–2019. Its ecological picture is asymmetric: grey squirrels displace reds mainly in low elevation and deciduous regions, while red squirrels persist predominantly in higher altitude and dense coniferous refugia. This suggests that squirrel spatial dynamics are strongly conditioned by habitat structure and by the competition coefficients linking native and invasive populations (Gramain, 2023).

3. Hibernation physiology, demography, and selection

For Arctic ground squirrels, the major physiological problem is the mechanism underlying interbout arousals during hibernation. The species hibernates for approximately 8 months, its body temperature falls from about $j$3 to as low as $j$4 via supercooling, and interbout arousals occur about 10–20 times per season. Two hypotheses are contrasted: the Torpor Arousal Clock Hypothesis, in which arousal intervals depend on a non-temperature-compensated circadian oscillator, and the Hourglass and Threshold Hypothesis, in which a hidden physiological quantity depletes to a threshold and is renewed during arousal. Sparse model selection from partially observed body-temperature data supports the Hourglass and Threshold Hypothesis and yields a two-dimensional nonlinear ODE with a hidden physiological driver:

$j$5

Here $j$6 is dimensionless body temperature, $j$7 is a hidden physiological state, and the model produces sustained limit cycles whose period is chiefly regulated by $j$8. During hibernation in the burrow the core model is autonomous, whereas other heterothermic species can be represented by adding circadian or circannual forcing terms (FitzGerald et al., 1 Apr 2025).

Columbian ground squirrels provide a complementary demographic perspective. Over a 32-year period, population density had negative effects on multiple demographic rates, including litter size at weaning, pup survival, proportion of breeding females, and adult female survival. Vegetation resources, measured through NDVI, had diverging effects. Mean NDVI had a positive effect on female yearling survival but negative effects on pup survival and litter size at weaning; NDVI slope had negative effects on survival rates but a positive effect on female mass gain over the reproductive season; and later NDVI phenology had positive effects on survival for pups, adult females, and yearlings. Neither population density nor vegetation affected population phenology or body condition in the following year. The resulting ecological picture is explicitly non-monotone: greater biomass or faster green-up did not uniformly improve outcomes, and later vegetation timing was favorable to survival for all ground squirrels (Tamian et al., 7 Jul 2025).

Selection analyses on Columbian ground squirrel emergence date from hibernation show that inference depends materially on the chosen fitness metric. Annual fitness was defined as

$j$9

where $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$0 is survival to the next breeding season and $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$1 is the number of offspring surviving to yearling. Lifetime reproductive success, matrix-based lifetime fitness from individual Leslie matrices, and mean lifetime annual fitness were then compared. All fitness metrics were highly correlated, but the estimated selection coefficients differed. In females, standardized directional coefficients were $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$2 for $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$3, $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$4 for $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$5, $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$6 for $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$7, and $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$8 for $y_{ij} \mid z_i, u_{ij} = \mathbbm{1}_{z_i > 0}\;\mathbbm{1}_{u_{ij} > 0}.$9; only the matrix-based and mean lifetime annual metrics detected stabilizing selection. In males, $z_i$0 and $z_i$1 yielded stronger directional selection for earlier emergence than either $z_i$2 or $z_i$3. A common misconception in long-term field analyses is that annual or lifetime reproductive success measures are interchangeable with lifetime matrix-based fitness; these results show that they can differ in both effect size and in whether stabilizing selection is detected at all (Dobson et al., 7 Jul 2025).

4. Locomotion, memory, and partially observed control

Squirrel behavior has also been abstracted as a coupled control-memory-verification problem. Empirically, fox squirrels adapt leap strategies to hidden mechanical properties of branches, learn to select more effective launch points, and use flexible landing maneuvers; their caching experiments show spatial chunking by nut species, and cache effort scales with food value and scarcity. Eastern gray squirrels recover their own caches after long delays and with altered landmarks, and they modify cache spacing and body orientation in the presence of observers. Field comparisons indicate similar capabilities in red squirrels. These behaviors are interpreted in minimal computational terms as rapid error correction under latent dynamics, memory for future control, and observer-dependent policy modulation (Armesto et al., 3 Apr 2026).

The resulting formalism introduces a hierarchical partially observed control model with embodied state $z_i$4, latent environmental state $z_i$5, structured episodic memory $z_i$6, observer-belief state $z_i$7, and task or permission state $z_i$8:

$z_i$9

Action selection proceeds through options, retrieval, and control,

$$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$0

with belief, memory, observer, and verifier updates,

$$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$1

The associated objective penalizes task inefficacy, latency, information leakage, and repair cost subject to competence constraints. Three hypotheses follow: fast local feedback plus predictive compensation improves robustness under hidden dynamics shifts; memory organized for future control improves delayed retrieval under cue conflict and load; and embedded verifiers with observer models reduce silent failure and information leakage while remaining vulnerable to misspecification. The paper is explicit that a proposer/executor/checker/adversary decomposition is a downstream conjecture rather than a direct claim warranted by squirrel evidence itself, and it imposes an inference ladder from empirical observation to minimal computational inference to AI design conjecture (Armesto et al., 3 Apr 2026).

5. Biomimetic robotics and squirrel-inspired engineering

Flying squirrels have inspired a family of agile aerial robots whose central design variable is a foldable membrane wing. One system integrates silicone foldable wing membranes into a quadrotor and coordinates them with thrust via the Thrust-Wing Coordination Control (TWCC) framework. The wing mass is 24 g out of a total mass of about 548 g, and aerodynamic forces from the wings are estimated with a physics-assisted recurrent neural network (paRNN) trained on real flight data and informed by flat-plate theory,

$$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$2

The learned model corrects the effective angle of attack and a deformation-related scaling factor to produce a real-time estimate of aerodynamic force. Experimentally, paRNN reduced aerodynamic-estimation RMSE by 24.4% to 70.4% relative to a vanilla RNN, and the complete flying squirrel drone achieved a 13.1% improvement in tracking performance, measured by RMSE, relative to a conventional wingless drone in high-speed outdoor obstacle-avoidance experiments (Lee et al., 13 Apr 2025).

A related platform uses controllable foldable wings fabricated from Ecoflex 0020 silicone membranes of thickness 0.3 mm and actuated by a 1-DoF crank-slider mechanism based on two coupled 4-bar linkages. Wing kinematics are described through the Freudenstein equation, wing area through the shoelace formula, and aerodynamic coefficients through flat-plate relations,

$$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$3

Because the elastic-wing aerodynamics are difficult to model explicitly, wing deployment is controlled by residual reinforcement learning initialized from human-demonstrated flight data through a Transformer-VAE and refined with PPO. The learned policy keeps the wings folded during acceleration and deploys them at deceleration onset. In fixed-time trajectories, terminal velocity reduction exceeded the nominal quadrotor baseline: for durations of 2.0 s, 2.3 s, and 2.6 s, the residual controller yielded 2.70 m/s, 3.82 m/s, and 4.95 m/s, versus 2.66 m/s, 3.76 m/s, and 4.80 m/s for the nominal quadrotor (Kang et al., 13 Apr 2025).

Tree-jumping squirrels have also motivated legged robotics. The 450 g robot Pinto was developed for jumping from the ground onto a vertical tree trunk, a maneuver explicitly identified as squirrel-like and rare in legged robots. Its actuation combines a twisted string mechanism with carbon fiber springs in a latched series-elastic architecture that can switch between series-elastic and parallel-elastic modes. The spring stores up to 2.53 J, and the latched parallel-elastic configuration increased jump energy by 11% to 1.42 J relative to rigid and series-elastic strategies. Pinto reached a center-of-mass apex height of 27 cm and perched at about 22 cm, using sprung 2-DoF arms with fish hook-like spined grippers to grasp bark asperities. A plausible implication is that squirrel-inspired design is especially productive where small-scale robots need both high-power ballistic phases and compliant, high-friction attachment phases (Xu et al., 2024).

6. Infrastructural interactions and non-biological extensions of the term

Squirrels also appear in infrastructure studies as a recurrent cause of distribution-level electric outages. Animal-related outages were a significant source of outages in Massachusetts from 2013 to 2018, accounting for about 7% of customers affected by all outages, and the broader literature cited in that study identifies squirrels as a leading cause or as second only to birds for overhead distribution outages in the United States. Earlier outage models, especially in Kansas, tied animal activity to squirrel seasonal and breeding patterns because species-specific activity estimates were unavailable. The study argues that such proxy-based models are limited: they do not capture regional variation, spatial heterogeneity, or multi-species structure. Its recommendation is to replace broad proxies with species-specific activity estimates, ideally through species distribution models and community-science data where available, while mitigation remains grounded in squirrel guards, clearance of overhanging vegetation, insulator and conductor modifications, and targeted patrols (Feng et al., 2021).

The word “Squirrels” also names a post-quantum signature scheme, and in that context it is unrelated to zoological classification. Squirrels is described as a GPV-style signature scheme on unstructured co-cyclic lattices with very large public keys and short signatures. For level 1 parameters, the public key is 681,780 bytes (665 kB) and the signature is approximately 1 KB. Standard verification checks the congruence

$$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$4

where $$z_i \mid \boldsymbol{\beta} \sim \text{Pois}(\lambda(\mathbf{x}_i)), \qquad \lambda(\mathbf{x}_i)=e^{\mathbf{x}_i' \boldsymbol{\beta}},$$5. A compressed-verification method replaces the large public key with a smaller verification key generated using random secret 31-bit primes, reducing Squirrels-I keys from 665 kB to 20.7 kB while also improving verification time. This is a terminological extension rather than a biological result, but it illustrates the breadth with which squirrel imagery now circulates across technical domains (Banegas et al., 3 Sep 2025).

Squirrel research, taken across these literatures, is therefore methodologically plural rather than taxonomically unified. The same label anchors latent-variable occupancy models, safety-niche competition systems, discrete invasion dynamics, nonlinear hibernation ODEs, long-term selection analyses, control-theoretic abstractions of locomotion and caching, morphing-wing drones, perch-capable robots, outage-risk modeling, and a conservative post-quantum signature scheme. This suggests that squirrels have become a recurrent comparative object precisely because they sit at the intersection of hidden-state inference, delayed consequences, spatial heterogeneity, and mechanically demanding behavior.