Bioinspiration & Biomimetics

Robotic herding offers a promising off-animal approach to controlling the spatial distribution of grazing livestock. Nevertheless, the habituation of cattle to herding stimuli remains a critical obstacle to its practical use. This study presents the first field investigation of behavioral responses of grazing cattle to herding by a quadruped robot. We compared the herding efficacy of a drone and a quadruped robot, and tracked daily changes in responsiveness during continuous herding. Trials were conducted in a 1.1-ha pasture with Japanese Black breeding cows from June to October 2023. The quadruped robot elicited stronger avoidance responses with shorter latencies than the drone, and herding was more successful with the robot. During consecutive daily herding with the robot, behavioral responsiveness declined progressively over the initial 5 days, at a markedly slower rate than previously reported for drone herding. Following a 24-day interruption, responsiveness partially increased, but declined rapidly again over the subsequent two days. These results indicate that the quadruped robot constitutes a more persistent aversive stimulus than a drone for grazing cattle, although habituation management strategies (such as diversifying stimuli or combining sensory modalities) will be necessary for sustained control of grazing distribution.

Known for their rapid water surface locomotion, whirligig beetles (Gyrinidae) are also fast divers. They can transition from the surface to underwater swimming in under 100 ms. Diving requires them to generate sufficient hydrodynamic torque to break the water surface and reorient the body. However, the mechanism by which whirligigs create the necessary pitch-down torque is unknown. Here, we show that their diving is achieved by bending the abdomen. The inertial hydrodynamic pressure on the flexed abdomen generates the torque necessary for surface breaking and body rotation. The torque is maximized at an abdominal angle of 45{degrees}, whereas for effective diving, whirligigs display a preferential abdominal angle of around 28{degrees}, closely aligning with the angle that maximizes the torque-to-drag ratio. This abdominal angle is optimized by balancing torque generation with the drag penalty. Our findings inspire a mechanism for surface-to-underwater transition, offering design principles for multimodal robots capable of locomotion across the air-water interface.

Collective motion emerges from local interactions among individuals, yet whether interaction rules inferred from trajectory data correspond to the mechanisms actually used by animals remains unresolved. Here, we address this question using an autonomous closed-loop robotic fish implementing a data-driven model of social interactions reconstructed from the schooling fish Hemigrammus rhodostomus. The robot continuously updated its behavior from real-time tracking of freely swimming fish while reproducing spontaneous locomotion, wall avoidance, and anisotropic attraction and alignment. We compared entirely biological groups, biohybrid groups containing one robotic fish, and numerical simulations using identical behavioral descriptors across isolated individuals, pairs, and groups of five fish. The robotic fish successfully integrated into natural schools and reproduced the principal signatures of collective coordination, providing the first direct causal validation of interaction rules reconstructed from behavioral trajectories. Biohybrid experiments showed that a robot responding only to its single most influential neighbor was sufficient to sustain natural collective coordination. By contrast, numerical simulations reproduced the behavior of biological groups most accurately when each fish interacted with its two most influential neighbors. This discrepancy identifies the contribution of hydrodynamic interactions, which remain available to living fish but are absent from the robotic controller, demonstrating that physical and behavioral interactions jointly shape collective organization. These findings establish closed-loop biohybrid robotics as a powerful framework for experimentally testing the mechanisms underlying collective animal behavior. SignificanceInferring the behavioral mechanisms underlying collective animal behavior from trajectory data alone cannot establish causality. We combined a data-driven model of fish social interactions with an autonomous closed-loop robotic fish that continuously interacted with freely swimming conspecifics. This biohybrid approach provides the first direct causal validation of interaction rules reconstructed from behavioral trajectories. Comparing biological groups, biohybrid groups, and numerical simulations further reveals that hydrodynamic interactions complement social interactions in shaping collective organization. While living fish require information from their two most influential neighbors to reproduce natural schools, a robotic fish lacking hydrodynamic feedback achieves comparable coordination by responding to only its single most influential neighbor, demonstrating the power of closed-loop biohybrid robotics for testing mechanisms of collective behavior.

Advances in camera technology and computer vision techniques have allowed researchers to track animals in 3D in ways which previously were difficult or impossible. Many such 3D tracking tools make use of multiple cameras, but unfamiliarity with the principles and technology involved can make it difficult to employ such techniques. In this protocol, we describe Braid, open-source software for live, multi-camera 3D tracking of insects. Using background-subtraction, Braid performs detection of objects without requiring the use of physical markers affixed to the insect. Braid constructs low-latency 3D position estimates using Kalman filtering and nearest neighbor data association. We document in detail the process of tracking freely flying bees within a flight arena using Braid. This protocol includes instructions on installation, configuration of cameras, setup, calibration, and operation. Within the system described here, we demonstrate that Braid can achieve position estimates accurate to <1 millimeter (within a 0.3 cubic meter volume). These factors make Braid suitable for tracking small, fast-flying animals like insects. Braid's low latency allows live tracking, removing the necessity to collect large video files and making it suitable for integration in closed loop systems such as virtual reality. Code is available at https://github.com/strawlab/strand-braid

Supernumerary robotic limbs (SLs) could extend human motor capabilities beyond the natural body, yet most demonstrations have been limited to simple, quasi-static tasks or direct teleoperation under visual control. Whether humans can safely and intuitively collaborate with autonomous SLs during complex, dynamic tasks remains unknown. We studied this question in a demanding assembly task requiring coordinated transport and handovers of a long object, which could not be completed by a single person. We developed a reconfigurable backpack platform mounting up to four high-payload robotic arms, together with safety-aware motion planning that generates human-compatible trajectories from task states without training data. Despite the SLs operating out of sight, participants adapted quickly, completed the task reliably, and reported low cognitive demand alongside high perceived safety and predictability. Kinematic and force analyses showed that control design governed both movement fluency and whole-body stability: faster limb motions yielded smoother interactions and shorter completion times, and a mirrored coordination strategy--in which one robotic limb counterbalanced the other--further shortened duration and reduced variability in left-right ground reaction force differences during co-manipulation. These results show that humans can integrate autonomous wearable limbs as out-of-sight partners in dynamic whole-body collaboration, and delineate behavioral and control principles underpinning fluency, stability, and usability in human augmentation.

Biomimicry has inspired the evolution of robotics toward greater autonomy, adaptability, and symbiosis with humans and dynamic environments. However, current robotic systems still face major challenges in recapitulating the high-efficiency decision-making capabilities of the human brain under complex and dynamic conditions. Here, we present Brainobot, a biohybrid robotic system that establishes a brain organoid controller as a high-level robotic decision-making layer for closed-loop embodiment. By leveraging brain organoid reservoir computing, Brainobot interacts with dynamic environments by receiving and processing sensory inputs and generating motor actions. As a proof-of-concept demonstration, Brainobot is implemented in a humanoid robotic system to perform real-world tasks, including object grasping and laser chasing. Interestingly, Brainobot exhibits unique features, including cross-task adaptivity, high computing efficiency, and low energy consumption. Thus, our approach may provide insights for advancing robotic embodiment and understanding biological decision-making.

Flight feather structures - shaft, barbs, and barbules - have transformed through evolution. Although size, shape and orientation of these structures likely influence the aerodynamic performance of the feathers, their impact remains unclear. Here we investigated the feather structures influence on flow and aerodynamic performance, using computational fluid dynamic modelling, by modifying the morphology of a section of a Jackdaw (Corvus monedula) flight feather. We show that the flow over the leading vane is a key factor, and that changing the barb-to-shaft angle from its natural state as well as the trailing vane barb angle, reduces performance, which could explain evolutionary changes from early Paraves species. All in all, our results suggest that current feather composition may be close to a local maximum for both aerodynamic efficiency and stability/flow predictability, suggesting evolution of feather structures have been driven by aerodynamic performance, which should have implications for designing bio-inspired technologies.

Developing robust methods to quantify how animals allocate time across behaviours is essential for understanding energy use, habitat requirements, and responses to environmental change. For cryptic, semi-aquatic mammals such as the platypus, direct observation is difficult, creating a reliance on remote biologging approaches that can reliably infer behaviour in the wild. However, aquatic environments can both smooth acceleration signals through hydrodynamic damping and introduce noise from water movement, turbulence, and drag, potentially obscuring behavioural differences of similar magnitudes. We tested whether progressively incorporating biomechanical and frequency-domain (FFT-derived) predictors improved behavioural classification in hydrodynamically challenging aquatic environments. Tri-axial accelerometers were deployed on four ex situ platypuses, with synchronised video observations used to validate behaviour. From the acceleration data, we derived three predictor classes of increasing complexity: summary statistics describing activity level, engineered biomechanical variables capturing posture and body orientation, and FFT-derived features describing movement rhythm. These predictors were progressively incorporated into Random Forest models to classify five behaviours: burrow resting, surface resting, grooming, travelling/foraging, and diving. Model performance improved with increasing predictor complexity, although gains were behaviour specific. FFT-derived features substantially improved classification of rhythmic behaviours such as diving and foraging, while engineered biomechanical predictors improved grooming detection. In contrast, resting behaviours, particularly surface resting, showed little improvement. Overall accuracy increased from [~]75% to [~]88% when frequency-domain features were included. Misclassification was greatest among behaviours with overlapping or low-amplitude signals, and cross-individual validation revealed reduced model generalisability, indicating that individual variation in movement patterns constrained transferability. Incorporating frequency-domain features substantially improved behavioural classification in platypuses, particularly for rhythmic behaviours such as diving and foraging. This study provides the first validated accelerometry-based behavioural classification framework for the species and highlights the importance of matching predictor selection to behavioural mechanics. More broadly, the approach offers a transferable framework for aquatic and semi-aquatic taxa.

Cruising speed is a key factor affecting prey-search efficiency and migration range in continuously swimming animals. Tunas and lamnid sharks (e.g., white sharks) have convergently evolved traits for high-speed cruising, including the ability to maintain slow-twitch, aerobic red muscle (RM) warmer than ambient water, known as RM endothermy. Despite their well-known high cruising speeds, kinematic features underlying their elevated speeds remain unclear. Swim speed is the product of tailbeat frequency (TBF; Hz) and stride length (SL, the absolute distance traveled per tailbeat; m). RM endothermy is expected to elevate TBF by enhancing muscle contraction performance. Furthermore, within RM-endothermic fishes, tunas and lamnid sharks may exhibit distinct kinematic features because of differences in caudal fin morphology and tailbeat amplitude. Here, we compiled kinematic parameters from 20 fish species, including five RM-endothermic species, measured in the wild using animal-borne sensors. Comparative analyses showed that, for a given body mass and water temperature, RM-endothermic fishes exhibited 1.9 times higher cruising speed and TBF than ectothermic fishes, while SL remained similar. Within RM-endothermic fishes, tunas exhibited 2.3 times higher TBF than similar-sized lamnid sharks, whereas lamnid sharks showed 1.7 times longer SL than similar-sized tunas. These results indicate that RM endothermy is generally associated with higher TBF, while significant kinematic differences remain between tunas and lamnid sharks. This divergence may be partly explained by the greater caudal fin area and tailbeat amplitude in lamnid sharks. It may also reflect contrasting skeletal types of teleosts and elasmobranchs, which potentially influence body stiffness and swimming kinematics.

Spiders arrest their fall by abseiling on dragline silk, varying silk release rate to avoid excessive strains that could break the thread. To understand how this is achieved, we develop a minimal dynamical model of abseiling and formulate a time-optimal control problem subject to physical limits on silk strain and dragline spinning rate. We find that the optimal trajectory includes an interval over which the strain is constant, resulting in uniform deceleration consistent with published experiments. To test the generality of this strategy, we then turn to a distinct extreme motor behavior seen in jumping spiders: using a dragline anchored at the takeoff site to correct their orientation while jumping at speeds exceeding 100 body lengths per second. We show that the same constant-strain strategy suffices to counteract the initial body rotation and ensure an upright landing without significant loss of range. Together, our theoretical framework suggests how strain regulation may serve as a strategy for controlling extreme movements such as abseiling seen in many spiders and may be specialized to tasks such as aerial braking and reorientation in jumping spiders.

Physical human-device interaction is central to many emerging technologies in neurorehabilitation and assistive robotics, but simulation-based research in this area remains fragmented across musculoskeletal models, assistive-device representations, task definitions, and controller-development workflows. This fragmentation limits the accessibility, reproducibility, and extensibility of studies on prostheses, exoskeletons, wearable rehabilitation devices, and related human-device systems. Here we introduce MyoAssist 1.0, an open-source framework for neuromechanical simulation of physical human-device interaction built within the MyoSuite ecosystem. MyoAssist organizes each simulation environment as a composed human-device-task system that combines compatible musculoskeletal, assistive-device, and task-scenario components through a shared composition pipeline. The current release includes 15 assistive-device models spanning gait assistance, upper-body support, manipulation, and seated mobility and supports compatible musculoskeletal models ranging from reduced lower-limb models to a 416-muscle full-body model. These human-device systems can be simulated within the broad task scenarios provided by MyoSuite, while MyoAssist adds locomotion-specific task scenarios with configurable terrain and target-velocity conditions for gait-assistive studies. MyoAssist also provides two complementary controller-development frameworks: a reinforcement-learning framework for training adaptive policies and a controller-optimization framework for tuning structured, interpretable human and device controllers. Both frameworks operate on the same simulation environments and provide standardized evaluation outputs for inspecting, comparing, reusing, and extending learned and structured control strategies. By integrating modular human models, assistive-device models, task scenarios, and training workflows under a shared open-source interface, MyoAssist aims to lower the barrier to reproducible simulation-based research and to support collaborative development of assistive technologies for neurorehabilitation and physical human-device interaction.

The Siamese fighting fish (Betta splendens) is well-known for its high levels of aggression and complex, stereotyped displays, making it an ideal model for testing deep learning-based tools for high-throughput, unbiased quantification of aggressive behaviour. A deep-learning object recognition model (BettaAI) was developed to detect the species' aggressive displays, including threat (opercula opening and fin distension), attack (bites) behaviours, as well as surface air breathing, a behaviour correlated with the frequency and intensity of aggression. Videos of males confronting their mirror image were acquired simultaneously with synchronised top and side cameras. Two separate models were trained, with some behaviours identified from top-view frames and others from side-view frames; outputs from both models were combined and post-processed for final quantification. BettaAI performed with high accuracy: F1 scores were 0.99 for the top-view model and 0.93 for the side-view model. Average Precision (AP) at 0.5 IoU exceeded 0.98 for all categories except bites, a rare behaviour in the dataset, which reached 0.865. Comparisons with a manually curated "ground-truth" dataset from videos of 10 males showed strong concordance between automated quantification and manual annotations by expert observers across all behaviours. A case study on 461 males reacting either to a mirror or to a video playback of a conspecific demonstrated the high-throughput and unbiased advantages of BettaAI, with threat displays being more frequent toward the mirror image and attack rates not differing between treatments. This work underscores the potential of BettaAI for automated, bias-free quantification of aggressive displays in B. splendens.

Marker-based optical motion capture remains the gold standard for kinematic analysis across biomechanical research. However, when tracking dynamic movements, severe occlusions often create missing data and noisy signals that require time-consuming manual cleaning. Existing automated methods usually rely on predefined skeleton models or need massive training datasets for machine learning solutions. This study presents the AMP: an automated marker based motion capture processing pipeline. The AMP is a fully analytical approach that uses hierarchical algorithms including 1) rigid-body constraint, 2) Kalman Filter state estimation and 3) kinematic pattern match to label markers and fill data gaps without manual human intervention. First, the algorithm identifies and removes mislabeling markers before any data scrubbing. During the data cleaning process, our hierarchical algorithm uses a translation-based step to track segments even when there are insufficient donor markers to form a full rigid body. To reconstruct missing marker trajectories, the pipeline initially uses Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) interpolation for short marker gaps. However, the PCHIP creates massive spatial errors during long occlusions, the system automatically switches to rigid-body tracking for gaps over 40 ms. Results showed the pipeline drastically reduced processing time compared to manual human expert methods by 73.5%. In addition, the AMP kept spatial errors within 10 mm of the manual ground truth. By using this combined hierarchical approach to match marker processing accuracy, the pipeline provides a fast, general-purpose tool that fully automates data cleaning and saves hours of manual effort.

Acoustic telemetry is a core methodology for collecting fine-scale movement data for aquatic animals. When telemetry receivers are set up in closely spaced arrays with overlapping detection ranges, the detection times of a tagged animal can be used to estimate its position and movement paths. In practice, estimating these paths can be challenging. Traditional time-difference-of-arrival methods generally provide positioning accuracy too poor for inferring fine-scale behaviours, while more robust state-space positioning models can be computationally intensive and practically infeasible to run on large datasets. Here, a novel telemetry positioning method is presented where time-of-arrival positioning is implemented as a state-space model within an extended Kalman filter. The resulting model, termed EK-TOA, provides closed-form solutions to track estimation. Simulated datasets of fish movement within a 2D telemetry array are used to verify the models performance and a real case study is provided where EK-TOA is utilized for the long-term tracking of a tagged fish. In comparison to currently available positioning models, EK-TOA provides a fast and accurate solution for tracking fine-scale movement behaviours of aquatic animals over long, continuous periods of time.

Planning husbandry tasks and experiments with Drosophila melanogaster requires converting a target date into development times that depend on the rearing temperature. This calculation needs to be done for each cross, genotype, and temperature, and the risk of error grows quickly. Available laboratory management tools let users register stocks, crosses, and track them, but they do not create schedules based on a clear, adjustable thermal model. To fill that gap, we developed DrosoTracker, a self-contained web application that works offline and predicts Drosophila development with a thermal summation model recalibrated through regression on data from Powsner (1935) (T0 = 11.78 {degrees}C, DD = 116.38 {degrees}C{middle dot}days, R{superscript 2} = 0.997). The model offers an optional two-level calibration driven by user observations. A wild-type strain first adjusts the model to the laboratorys own conditions. Then each genotype is calibrated against that reference using a random-effects shrinkage estimator that accounts for measurement error and between-batch variability. The model creates schedules for husbandry tasks, evaluates adult cohort survival with the Kaplan-Meier estimator and the log-rank test, and calculates sample size for lifespan studies using Schoenfelds formula. The quantitative components were checked against independent references, including Rs survival package and manual calculations. Ongoing work is focused on validating the calibrated model using cohorts specifically bred for this purpose. DrosoTracker runs entirely in the browser, stores data locally, and is available in English and Spanish.

Animal pose estimation typically requires manual annotation of body parts to train neural networks for each new recording system or environmental condition. Here we use a general-purpose vision-language model (VLM) to generate annotations in multi-animal recordings of fruit flies (Drosophila melanogaster). Initial annotations follow written anatomical descriptions and are reviewed by an independent VLM session using neighboring video frames. On 120 held-out fly images, AI annotations had a median distance of 2.28 pixels from human labels, with 84.0% agreement within eight pixels, averaged across landmark types. The comparison included 10 landmarks, with head and eyes analyzed separately because of systematic differences in placement. Review improved agreement primarily at the feet, with the largest remaining discrepancies at positions with limited visual evidence. Visual inspection revealed cases in which the AI annotations were more anatomically accurate than those of an experienced human annotator, particularly at the feet. A convolutional network trained using SLEAP on 300 AI-annotated fly images showed similar agreement with human labels and closely reproduced the AI annotations on held-out images, with a median distance of 1.56 pixels. We also applied the procedure to recordings from another laboratory, using new AI annotations to train a separate pose network. Although VLM annotation required minutes per fly, the trained networks enabled pose estimation throughout more than 700,000 frames. These results demonstrate that VLM annotations can support fast animal pose estimation without manually labeled training images.

The spring-loaded inverted pendulum (SLIP) model has effectively explained center of mass (CoM) and ground reaction force (GRF) dynamics during human gait despite its simplicity, describing both walking and running within the same mechanical principles. A recent study extended the explanatory capability of the SLIP model to joint dynamics by reproducing ankle joint dynamics during walking. However, during running, the model did not capture the empirical ankle torque profile, suggesting that such mechanical unification may not hold at the joint level. Because the previous model represented the body as a single point mass, it neglected leg inertia, which may become important during running as rapid motion generates substantial angular momentum. To address this limitation, we propose an extended two-mass model that separates the whole-body mass into the swing-leg and the remaining body mass, thereby explicitly incorporating swing-leg inertia into stance-leg joint dynamics. The swing leg is modeled using a linear spring and a torsional spring at the hip joint for its radial and rotational motion. Through inertial coupling, swing-leg dynamics influence the stance-leg ankle torque. Consequently, the proposed model reproduces the symmetric, single-peaked ankle torque profile observed experimentally while preserving the CoM and GRF of the previous model. These findings suggest that the discrepancy of the previous model during running may be explained by the omission of swing-leg inertia. By accounting for the gait-mode-dependent contribution of swing-leg inertia, the proposed model provides a unified mechanical framework for extending simple gait models from CoM-level behavior to joint dynamics.

Gliding enables mammals to forage and escape from predators by moving between discontinuous forests. Its benefits depend not only on glide distance but also on the ability to decelerate and land safely. This study developed a theoretical framework linking glide distance, gliding velocity, aerodynamic braking, body mass, and braking distance. Twenty-two representative distance-velocity observations from eight studies and five species were compiled. Among distance-velocity models, log-distance and saturated with V0 models received nearly equivalent support. Both models predicted increasing velocity with glide distance, with the rate of increase declining at longer distances. For a 1 kg animal undergoing a 60% reduction in velocity, predicted kinetic energy remaining immediately before contact increased from 4.80-5.14J at 20 m to 8.13-8.90 J at 80 m. This velocity reduction corresponded to a dissipation of 84% of approach kinetic energy before contact. Over a braking distance of 1 m, the required mean deceleration increased from 2.57-2.75 g at 20 m to 4.35-4.77 g at 80 m. At 80 m, shortening the braking distance from 4 to 0.5 m increased the required deceleration from 1.09-1.19 to 8.70-9.53 g. These results indicate that the absolute energetic and deceleration requirements of landing increase with glide distance, even when velocity is close to an asymptote. These results provide a quantitative basis for considering aerodynamic braking and landing requirements alongside conventional measures of glide performance. Summary statementModels quantify how glide distance, aerodynamic braking, and braking distance affect pre-contact kinetic energy and deceleration requirements in gliding mammals.

Musculoskeletal simulations are widely used to estimate joint loading, yet most musculoskeletal models estimate knee contact forces as resultant forces or as normal medial and lateral contact forces, without resolving pressure distributions across the articular surfaces. This paper presents a smoothed mesh-based knee contact pressure model that computes continuously differentiable tibiofemoral contact pressures, enabling its direct integration into full-body movement simulations. Built on an elastic foundation formulation, the model introduces smooth approximations ensuring that all contact functions and their derivatives remain continuous throughout the simulation. A systematic sensitivity analysis was performed across five key parameters: mesh resolution, joint damping and three smoothing parameters. Tracking simulations across eight gait trials demonstrated that the nominal configuration achieved mean RMSE values for medial and lateral knee contact forces of 51.6 N and 75.2 N, respectively, with a mean RMSE for joint angles of 1.45{degrees} and (r = 0.97), converging in less than three hours on a standard computer. Mesh resolution was identified as the dominant factor that affected both accuracy and convergence, while damping variations had negligible influence. As a proof of concept, the model was also incorporated into predictive simulations, demonstrating that increasing the weight on the contact pressure term in the cost function leads to reduced tibiofemoral loading, particularly in the lateral compartment. The proposed formulation provides a computationally efficient and numerically robust framework for simulating knee contact mechanics within full-body musculoskeletal models.

Interception of moving objects is a universal pattern of animal behavior with many industrial applications. Inspired by the behavior of the characid fish Brycon guatemalensis, which catches fruits falling onto the surface of a river, we study one-shot (i.e., open-loop) interceptions using both stochastic optimal control theory and reinforcement learning. In the first instance, an agent actuates a control (wait time before launch and heading angle) to meet its target, given a noisy estimate of the target initial state. In a minimal setting, where launch velocity and actuated control are constants, we show that the optimal wait time is attenuated by noise in both the target measurement and control actuation, as well as agent energy conservation. We then derive an upper bound for the heading angle variance which scales as the ratio of the non-vanishing optimal interception error to the initial planar separation. Complementing our control framework, our deep reinforcement learning approach shows that the neural network representing the fish first learns the value function for the interception, then learns how to aim, and finally learns the optimal wait time. Beyond the specific problem at hand, our results may also have applications to problems with little time for feedback or corrections such as perching, landing, and docking.