In situ swimming behavior of the Mariana snailfish Pseudoliparis swirei Authors: Yizi Chen, Alexandru S Barcan, Jiushuang Zhang, Jun Li, Lisheng He, Yong Wang Published: 2026 ABSTRACT AbstractThe hadal zone, defined as ocean depths below 6000 m, is one of Earth’s most extreme environments. The Mariana snailfish Pseudoliparis swirei is among the deepest known fishes, yet its natural behavior remains largely unknown. Here we quantify, for the first time, the three-dimensional swimming behavior of this species in its native habitat by analyzing video collected by a baited deep-sea lander at nearly 7000 m depth. Using an automated computer vision workflow that detects, tracks and infers visual depth from single camera footage, we reconstructed full trajectories for individual fish using 868 manually annotated instances. The results show that P. swirei exhibits a slow routine swimming speed of 0.16-0.18 m s⁻1 (0.62-0.80 body lengths s⁻1, mean ≈ 0.71 BL s⁻1), with low variability across individuals and only short acceleration events reaching up to ~ 0.50 m s⁻1. By combining reconstructed movement paths with a simple representation of near bottom flow, we estimate that the fish operates within a modest spatial range around the bait. A mechanistic advection-diffusion framework further indicates that P. swirei first detects bait odor at a distance of approximately 350 m. These findings provide the first quantitative view of locomotion in a hadal vertebrate, revealing gentle, steady gaits shaped by environmental constraints.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-48409-2.Keywords: Snailfish, Hadal trench, Swimming speed, Deep-sea fish, BiomechanicsSubject terms: Ecology, Ecology, Ocean sciences, Zoology Introduction The hadal zone, encompassing ocean depths greater than 6,000 m, represents one of Earth’s most extreme habitats1. This environment is defined by permanent darkness, near freezing temperatures, immense hydrostatic pressures, and limited nutrient availability2. Organisms inhabiting these depths experience significant energetic constraints, necessitating specialized adaptations for survival3. Among vertebrates, fishes encounter particularly stringent challenges, and their distribution is markedly restricted at extreme depths. Beyond approximately 8,200 m, snailfishes belonging to the family Liparidae are currently the only known vertebrates inhabiting these depths, and they dominate the hadal ichthyofauna4. The Mariana snailfish, Pseudoliparis swirei, first described in 2017 from depths between approximately 6,900 and 8,300 m, is among the deepest known vertebrates inhabiting hadal trench ecosystems5, although recent observations indicate that Pseudoliparis belyaevi may occur at slightly greater depths6. P. swirei has evolved extraordinary physiological and anatomical adaptations that allow survival under conditions of extreme hydrostatic pressure, low temperature, and sparse food availability4,5,7. P. swirei exhibits several distinctive morphological traits consistent with its specialized hadal lifestyle4,8–11. The fish possesses a translucent and scaleless body, significantly reducing structural density5. Its skeletal system exhibits marked reduction in ossification, particularly evident in its flexible and incomplete skull, adaptations presumably facilitating tolerance to immense pressures8. Indeed, the only extensively ossified elements observed within this species are the otoliths and teeth, indicating selective structural reinforcement10. The fragility of P. swirei makes specimens difficult to preserve when brought to the surface, as the subdermal gelatinous tissues can relax and become damaged at room temperature or during preservation6,12. Collectively, these morphological characteristics strongly suggest a metabolism adapted to minimize energy consumption, congruent with broader trends observed among deep-sea fishes13–16. Traditional perspectives on abyssal and hadal ichthyofauna characterize snailfish populations as slow-moving benthic predators or opportunistic scavengers, reflecting energy conservation strategies imposed by their nutrient-poor environment12,17. Species such as grenadiers (e.g., Coryphaenoides armatus) illustrate this paradigm clearly, exhibiting notably low metabolic rates and slow, energy-conserving movements primarily aimed at locating infrequent food sources, typically carrion falls7. Similarly, the bathyal gadiform fish, blue hake (Antimora rostrata), demonstrates reduced basal metabolic activity coupled with limited muscular power, yet paradoxically retains short bursts of swimming speed comparable to shallow-water fish at equivalent temperatures18,19. This observation suggests that, while depth-related factors constrain routine activity, fundamental locomotor capacities remain intact. These behaviors, shaped by the combined pressures of cold, darkness, high hydrostatic stress, and resource scarcity, emphasize energy conservation and restrained movement. Historically, abyssal plains were hypothesized to favor passive sit-and-wait predation strategies, although modern telemetry and baited lander studies have revised this view20,21. Tagged grenadiers, for instance, actively seek out and intercept food sources rather than adopting entirely stationary foraging strategies, yet they still maintain an overall strategy of minimizing energetic expenditures7. Hadal snailfishes, including P. swirei, also appear to forage actively, but available observations suggest that they rely more on predation on abundant small benthic prey, particularly amphipods, than on scavenging large carrion falls5,22. The snailfish predominantly prey upon abundant amphipods, such as Hirondellea species, employing active suction feeding rather than passively awaiting large carrion deposits4,17. Trenches, including the Mariana Trench, typically exhibit weak currents23,24. Long-term moorings in the Challenger Deep record near-bottom currents averaging less than 0.01 m s⁻1, occasionally punctuated by brief tidal-driven bursts approaching 0.08 m s⁻1. Such sluggish hydrodynamic conditions severely limit passive drift capabilities in the hadal environment24, potentially constraining the spatial extent over which organisms encounter food resources. In the unique ecological context of the Mariana Trench, P. swirei effectively functions as an apex predator despite its fragile morphological appearance4,22. Despite the prominence of snailfishes in hadal ecosystems, virtually nothing is known about their natural behavior in situ. The extreme logistics of observing animals at abyssal and hadal depths (remote location, high pressure, no light) have historically limited direct behavioral studies14. Here, we address this knowledge gap by quantifying the locomotion of Pseudoliparis swirei directly in its natural hadal environment. Using video footage recorded by a baited deep-sea lander deployed at 6,670 m depth in the Mariana Trench, we developed a computer-vision workflow capable of detecting, tracking and reconstructing the three-dimensional movement trajectories of individual fish from single-camera recordings. The analysis was based on three independent encounters with individual snailfish observed in the footage. This approach allowed us to estimate routine and burst swimming speeds, characterize movement patterns near the seafloor and examine how locomotion is shaped by the physical and ecological constraints of the hadal environment. By combining trajectory reconstruction with a simple representation of near-bottom flow and odor plume dispersion, we further estimate the spatial scale over which P. swirei detects and approaches food sources. Together, these analyses provide a quantitative description of locomotion in one of the deepest known vertebrate and offer new insight into behavioral strategies that enable fishes to persist under the extreme conditions of the hadal zone. Results To quantify in situ swimming behaviour of P. swirei we applied the automated analysis pipeline to video recordings obtained from a baited deep sea lander in the Mariana Trench. The segmentation network identified and tracked the fish in each frame that contained an individual, while the Depth Anything ViT B model produced a depth map for every image. Combining these outputs with the camera calibration allowed us to reconstruct three dimensional trajectories and compute instantaneous swimming speeds for three separate encounters with P. swirei. We then used these trajectories to describe routine and peak speeds and, in a second step, to obtain an approximate estimate of the spatial range over which P. swirei can move around the lander. Automated detection and depth estimation of P. swirei Representative detection and depth outputs are shown in Fig. 1. In a typical frame the network correctly delineated the fish body and assigned a unique track identifier, even when the animal was close to the colourful lander frame or the bait pile (Fig. 1a). The corresponding depth map (Fig. 1b) preserved the overall layout of the scene. Objects closest to the camera, such as the lander frame and bait, appear as shallow regions, whereas the fish body and background sediment are mapped to greater relative depth. The red dashed box highlights the region used to derive the fish depth, and the inset panel summarises frame index and detection information exported by the pipeline. Qualitative inspection across many such frames indicated that the combination of segmentation and monocular depth inference was stable in the low light, high backscatter conditions encountered at hadal depth. Example frame showing detection, segmentation and depth estimation of P. swirei in lander footage. Panel A, shows the original RGB frame from the Challenger Deep lander sequence with the segmentation mask of P. swirei shown in red and the track identifier. In Panel B we envision the corresponding depth map generated by the Depth Anything ViT B model. The dashed red box marks the fish region and the inset panel reports the frame index and detection information. To illustrate how consistently the network detects the fish across the sequence, we compiled sixteen example frames into a montage (Fig. 2). Example detection and segmentation of P. swirei in training images used for model development. Sixteen representative frames showing predicted bounding boxes and segmentation masks (blue). The examples include variations in distance, cropping, and viewing angle, reflecting the diversity of the training dataset and illustrating robust detection under different visual conditions. These frames span a range of distances from the camera, different body orientations and both open water and near-structure backgrounds. In all examples the model draws a tight bounding box around the fish and recovers a plausible segmentation mask, including the head and caudal region when they are visible. Failures were rare and typically associated with extreme motion blur or heavy occlusion when the fish passed very close to the camera housing. Together with the training and validation curves (Supplementary Fig. S1) that indicate high precision and recall on a held-out dataset, these examples demonstrate that the network provides reliable localisation of P. swirei throughout the analysed footage. The dataset comprised 868 manually annotated instances of P. swirei (Supplementary Fig. S2). Representative examples from the training dataset used for model development are shown in Fig. 2; these images are not sequential frames from the single-camera lander recordings used for behavioral analysis. Swimming speed profiles of P. swirei We quantified three-dimensional swimming speed for three P. swirei individuals recorded by the baited lander. For each fish we computed frame by frame three-dimensional speed from the reconstructed trajectory and then applied a 21 frame moving average to obtain a smoothed speed profile. The three time series are shown in Fig. 3a-c, with raw speeds in light red and smoothed speeds in blue, green, and orange, and a summary of mean and maximum smoothed speeds for each individual is given in Fig. 3d. Instantaneous three dimensional swimming speed of three P. swirei individuals. (a-c) Frame by frame 3D swimming speeds and 21 frame moving averages (blue, green, and orange) for fish 1–3. All three records show similar patterns of routine movement with repeated short accelerations. (d) Summary of mean and maximum smoothed speeds for each fish. Mean smoothed speeds were 159.3178.1 mm s⁻1 for fish 1–3, corresponding to 0.62–0.80 body lengths s⁻1 (mean ≈ 0.71 BL s⁻1), consistent with standard metrics used in fish swimming studies14,25, and maximum smoothed speeds were 309.9, 504.6, and 429.3 mm s⁻1 for fish 1–3. Swimming speeds were additionally normalized to body lengths per second (BL s⁻1); details of body length estimation and calculation are provided in the Supplementary methods: Body length estimation and BL s⁻1 calculation. Although the recordings differed in duration, all three individuals showed very similar swimming patterns (Fig. 3a-c). Each time series is dominated by a band of routine speeds around 0.16 to 0.18 m s⁻1, with repeated short accelerations on top of this baseline. High instantaneous peaks were attributed to tracking and depth-estimation noise and were excluded from interpretation; therefore, analysis focuses on smoothed speeds. The summary panel (Fig. 3d and caption) shows that mean and maximum smoothed speeds are closely matched among individuals, indicating that our estimates are consistent across independent recordings. Very large instantaneous spikes above 1000 mm s⁻1 were rare and were treated as artefacts by the anomaly detection step. Estimated dispersal radius of movement Based on the plume-response model, P. swirei is estimated to detect the bait odor at distances of approximately 350 m from the lander. In the advection–diffusion and fish response framework described in the Methods, this distance is defined as the point at which the modeled plume extent and the fish’s subsequent displacement are spatially consistent under the observed total elapsed time, rather than implying equality between plume travel time and fish response time. Based on the sequence specific parameters and precise timestamps, the model produced detection radii of 347.2, 348.6 and 356.8 m for the three analysed sequences, and random walk displacements of 5.7, 6.2 and 6.6 m. These values correspond to initial position intervals of 341 to 364 m at the time of bait deployment. The random walk term is therefore very small compared with the detection radius, so the fish position at first odor perception is effectively the same as its position at bait deployment. Taken together, these results indicate that an odor detection distance of approximately 350 m is a model-based estimate for the range over which P. swirei first engages with bait plumes on the trench floor. A schematic three-dimensional representation of this geometry, including lander position, bait diffusion area, current direction and an example horizontal trajectory, is shown in Fig. 4. Spatial structure of odor detection and movement in P. swirei. (a) Three-dimensional representation of the same spatial structure, illustrating the geometry of the odor field and current relative to fish movement. (b) Two-dimensional schematic showing reconstructed swimming trajectories of P. swirei (blue, orange and red lines) relative to the baited lander (red dot), together with the modeled odor diffusion field at the time of first detection (green region), the estimated random-motion envelope during plume propagation (yellow boundary), and the prevailing near-bottom current direction (blue arrow). Discussion Estimating the in-situ swimming speed of P. swirei in three dimensions provides new quantitative insight into the locomotion of a hadal vertebrate in its natural environment. Deep-ocean trenches are among the most challenging environments on Earth, defined by crushing pressure, perpetual darkness, near-freezing temperatures, and extreme remoteness, which make direct observation exceedingly difficult. Against this backdrop, our application of a monocular video tracking pipeline using footage from a deep-sea lander represents a significant technical and biological advance. Here, we quantify how one of the deepest known fishes moves in its natural habitat. The Mariana snailfish was found to cruise at a gentle, unhurried pace, with individuals maintaining routine speeds on the order of only a few centimeters per second (≈0.6–0.8 body lengths s⁻1). This slow locomotion aligns with the well-documented pattern of consistently low swimming speeds in deep-sea fishes, irrespective of phylogenetic affinity14,15,26. In the hadal realm there are essentially no fast visual predators, so selection does not favor rapid escape responses27,28. Under such conditions, a slow cruising speed is not a disadvantage but can be advantageous for survival, because it reduces energy expenditure without greatly increasing predation risk29,30. Indeed, large benthic fishes such as grenadiers (Coryphaenoides spp.) and Antimora rostrata rarely exceed ~ 0.1–0.2 m s⁻1 even when foraging25. Although these species differ substantially in body size from P. swirei, their locomotion similarly reflects the generally slow movement patterns typical of deep-demersal fishes. All three recorded snailfish exhibited similar velocity profiles, suggesting broadly consistent locomotor behavior among the observed individuals. The slow speeds recorded for P. swirei also provide insight into how this species forages in the hadal environment. Rather than chasing agile prey, hadal snailfish are likely opportunistic predators that feed on small benthic invertebrates such as amphipods encountered on the trench floor4,17. Deep trenches host a variety of benthic invertebrates, including amphipods and other detritivores, which may aggregate around organic material on the seafloor31. P. swirei appears to exploit these resources by swimming slowly and remaining close to the seafloor while searching for prey. All individuals in our study maintained a head-down, near-bottom posture, a behavior broadly consistent with chemosensory foraging described in other deep-sea fishes such as grenadiers32. The snailfish’s slow locomotion likely enhances its ability to detect food via smell or touch in the dark, while also preventing unnecessary disturbance of prey. When prey is encountered, P. swirei likely captures it using suction feeding, aided by its large gape and pharyngeal jaws used to process small crustaceans5,16,33. Such feeding does not require sustained speed, because amphipods and other hadal crustaceans are relatively slow or clumped around carrion34, so the snailfish can capture them with minimal energetic expenditure. This strategy aligns with the ecological role of snailfishes as dominant predatory fishes in many hadal trench ecosystems8. Unlike fast-swimming pelagic predators, P. swirei relies on efficient, low-energy foraging on relatively slow benthic invertebrates. The consistency of slow cruising across all observed individuals suggests that this feeding mode is species-typical: P. swirei patrols methodically and waits for prey to come within range, rather than actively pursuing anything over long distances. Such behavior is well-suited to a deep-sea predator that must maximize energy gained from every meal while living in an environment where food availability can be limited35. The measured locomotor traits of P. swirei are furthermore congruent with its unique morphological and physiological adaptations as a hadal fish36. This species’ anatomy effectively trades fast, powerful movement for buoyancy and pressure tolerance8,37,38. Like many benthic fishes, the hadal snailfish lacks a swim bladder12. Instead, it achieves near-neutral buoyancy through low-density gelatinous tissues and a lightly ossified skeleton17. Our findings of slow, neutrally buoyant gliding comport with this design. The extensive subdermal gelatinous layer in P. swirei likely allows it to float with minimal effort and may even streamline the body39. In P. swirei, the gelatinous tissue (which is > 95% water) may function as a fairing that smooths water flow over the body, enabling the fish to sustain movement using very gentle fin beats38,39. The snailfish’s primary locomotor appendages are its large pectoral fins, which it undulates in a wing-like fashion to propel itself; these fins are broad but flexible (“fragile” in texture), well-suited for pushing against water at low speeds but ill-suited for powerful thrusts40. The tail of P. swirei is small and continuous with the dorsal and anal fins, forming a low-aspect-ratio fin, a shape associated with maneuverability and efficiency rather than rapid propulsion25. This is a common trait in deep-sea fishes that rely on gelatinous tissues. Growth is inexpensive and buoyancy is improved, at the expense of muscular power and speed8,39. P. swirei perfectly embodies this trade-off, using a morphology that maximizes survival in the hadal niche even if it means accepting a very unhurried lifestyle. Comparisons with other taxa reinforce that P. swirei’s locomotion is a product of its environment and evolutionary history. Shallow-water snailfishes (family Liparidae) are generally small, benthic fishes that do not exhibit rapid swimming, and many possess suction discs used to attach to substrates, consistent with a lifestyle characterized by benthic association and limited movement40. The hadal snailfish’s ancestors diverged from shallow liparids ~ 20 million years ago8, and in adapting to deeper waters they have further reduced their locomotor capacity in favor of other survival strategies. In the abyssal zones (4,000–6,000 m) just above P. swirei’s range, the fish fauna shifts to larger scavengers like rattails (grenadiers, Macrouridae) and cutthroat eels (Synaphobranchidae)27,41,42.These fishes also move slowly, consistent with cold, high-pressure conditions: for example, the routine swimming speed of an abyssal grenadier is on the order of 0.05–0.1 m s^−1, with an upper limit around 0.2 m s^−125. Long-term baited camera studies have documented that such scavengers move slowly while searching the seafloor32. Our in situ observations place P. swirei squarely in this same low-velocity category, extending the trend into the hadal zone. Deep-demersal fish communities are typically dominated by species exhibiting relatively sluggish locomotion. Across diverse taxonomic groups, most deep demersal fishes move at consistently slow speeds, punctuated only by infrequent bursts14. This appears to be a general adaptation to the dual challenges of low temperature (which slows muscle function) and limited food (which favors energy economy) in the deep ocean8,14. In the hadal trenches, water temperatures hover just above freezing43 and prey yields relatively little energy per individual item. Under these conditions, it is not surprising that P. swirei exhibits slow swimming speeds, consistent with the energy-limited conditions of the hadal environment. Moreover, unlike the more migratory grenadiers or eels, snailfish are small-bodied and benthic. P. swirei remained close to the seafloor and moved slowly in the recorded sequences, consistent with a low-energy, near-bottom feeding strategy. The evolution of P. swirei and its hadal relatives can thus be considered as the culmination of a trajectory toward slower locomotion with increasing depth, a trajectory mirrored by other deep-sea lineages that replace speed and endurance with efficiency and sufficiency14,25. The ecological and evolutionary implications of P. swirei’s slow swimming are significant. Energetically, maintaining a low cruising speed helps ensure that the snailfish’s energy intake (from eating amphipods and other meiobenthos) can meet or exceed its expenditures in the impoverished hadal environment8. The hadal snailfish likely has a low basal metabolic rate to begin with, consistent with the general decline in metabolic rates observed with depth13. By seldom accelerating beyond a glide, the snailfish conserves precious calories and can allocate more energy to growth, reproduction, and sustaining vital functions under pressure. This strategy may also underlie the species’ surprisingly high abundance in trenches8, by keeping metabolic demands low, many individuals can be supported by the limited food supply. On the other hand, the limited locomotor capacity constrains how far a snailfish can travel in search of resources. If local conditions deteriorate, a snailfish cannot quickly migrate to a new site; instead, it must endure or slowly relocate over time. This lifestyle has consequences for dispersal and gene flow. The observed swimming speeds imply that adult P. swirei are incapable of long-distance movement across the vast abyssal plains separating trench systems. Yet, intriguingly, genetic studies show that hadal snailfish populations in different trenches such as Mariana, Japan and Kermadec are closely related, indicating that some dispersal has occurred between trenches within the past million years44. The most plausible explanation is that dispersal happens during early life stages. Perhaps buoyant eggs or larvae are transported by deep currents to colonize new trenches, since the adults themselves are unlikely to swim out of a trench. This dynamic of local sedentary behavior and occasional long-range dispersal parallels other deep-sea fauna (e.g. amphipods and holothurians) where population connectivity relies on passive transport rather than active movement. Finally, P. swirei’s locomotor strategy influences its habitat usage. These snailfish are essentially hadal endemics. They thrive from ~ 6,000 m down to the trench bottoms (~ 8,000 + m) but are absent from higher slope depths where more mobile competitors exist8. Our diffusion-response modelling (Supplementary Method M1) indicates that P. swirei first detected the bait odor at an approximate distance of 350 m from the lander, consistent with a foraging strategy adapted to sparse and patchy resources in the hadal zone. This scale is illustrated in Fig. 4, which shows the reconstructed three dimensional trajectory of P. swirei relative to the baited lander, the modelled odor diffusion field and the prevailing near bottom current direction. The trajectory follows a downstream approach pattern, suggesting that P. swirei may detect and track odor plumes over distances on the order of several hundred meters while navigating within weak but persistent near bottom flow. This interpretation aligns with recent observations of a three layer circulation structure in the Challenger Deep, including westward transport of lower circumpolar deep water, mid depth cyclonic rotation and bottom layer anticyclonic flow24. These hydrodynamic conditions could facilitate downstream dispersal of chemical cues and support energetically efficient foraging trajectories for benthic predators such as P. swirei. Future studies with larger video datasets will enable quantitative analyses of population density, diel patterns, and behavioural variation across depths and trenches. Beyond the biological insights, our study highlights the value of new video-based methods for deep-sea research. The ability to reconstruct three-dimensional movements from a single-camera video, validated by repeat observations, opens a pathway to quantifying behavior in environments where direct human observation or experimental manipulation is virtually impossible. Traditionally, nearly all knowledge of hadal fish came from recovered specimens or brief two-dimensional video encounters, leaving many questions about their natural behavior unanswered44. In situ observations using remotely operated vehicles (ROVs) provide valuable insights into the behaviour and ecology of deep-sea fishes42. By using a monocular vision pipeline and careful calibration, we show that even in the deepest ocean it is possible to measure how an animal moves, including its speed, its trajectory, and its spatial use of the habitat, without removing it from its environment. Such measurements may help bridge behavioral observations with physiology, ecology and evolution in species that cannot be studied through traditional laboratory or ship-based methods, complementing broader efforts to understand fish biology through integrated experimental systems22,45,46. Furthermore, this technological achievement provides a more holistic view of P. swirei in situ and can be applied to other species as well. For example, similar video tracking could be used to study diel activity patterns, social interactions, or avoidance responses in hadal organisms, none of which have been quantitatively described before. The incorporation of modern computer vision and machine learning tools will further enhance this field: already, researchers have developed deep-learning models to automatically detect and track hadal snailfish in video footage47. Such advances will dramatically increase the volume of behavioral data collectable from deep-sea expeditions, allowing scientists to move from anecdotal observations to robust datasets on locomotion and ecology in the abyss. Ultimately, connecting these behavioral observations to physiological and genetic data will deepen our understanding of how life functions under extreme pressure. For instance, knowing P. swirei’s routine speed and activity level in situ provides important context for its metabolic enzyme profiles and genome adaptations (e.g. those affecting muscle performance or energy use)8,44. It also informs models of hadal ecosystem energetics, and how much food a top predator requires and how far it might roam, which in turn affect nutrient cycling and prey population dynamics in the trench. While the present analysis is based on only three observed individuals, it provides an initial quantitative characterization of locomotor behavior in P. swirei under natural hadal conditions. In summary, the present study not only elucidates the locomotor behavior of the Mariana snailfish from a biological perspective, but also demonstrates a powerful approach to studying deep-sea life on its own terms. The remarkable consistency and slowness of P. swirei’s swimming likely reflect both the physiological constraints of the hadal environment and locomotor traits common within the family Liparidae. Slow swimming behaviour is typical of many members of Liparidae across a wide range of depths, suggesting that the locomotor strategy observed in P. swirei arises from both family-level morphology and environmental conditions. As we continue to refine in situ measurement techniques, we can look forward to a more complete picture of how hadal animals move, feed, and live, knowledge that is essential for understanding life at the very limits of the biosphere. Materials and methods Video acquisition During the expedition of R/V Tansuo 01 on July 13, 2016, video footage was recorded by a baited lander deployed at 6,670 m depth in the Mariana Trench (141°52.56′E, 10°56.00′N). The elapsed time from lander deployment to the first appearance of snailfish was 10 h 34 min 15 s, 10 h 40 min 28 s, and 10 h 52 min 1 s. The duration of the three video observations was 1 min 35 s, 8 min 5 s, and 5 min 49 s. The camera recorded at a frame rate of 30 fps with a resolution of 1920 × 1080 pixels. All observations were non-invasive, and no live animals were handled during this study. Analysis pipeline We developed an automated analysis pipeline to process underwater video and obtain swimming speeds of P. swirei. Throughout the manuscript, “water depth” refers to the bathymetric depth of the study site (approximately 6670 m), whereas “camera depth” or “relative depth” denotes the distance between the fish and the camera, estimated via monocular image analysis. The pipeline performs object detection and segmentation, tracking, monocular depth estimation (see below), two and three dimensional velocity calculation, anomaly detection, and visualization with result export in a single pass. For each video run, the workflow was as follows. Segmentation detection. The target fish was detected in each frame, followed by generation of a segmentation mask, bounding box and center point. Segmentation detection. The target fish was detected in each frame, followed by generation of a segmentation mask, bounding box and center point. Depth estimation. The Depth Anything model was used to produce a pixel level depth map for every frame. Depth estimation. The Depth Anything model was used to produce a pixel level depth map for every frame. Depth sampling. Within the fish mask we applied a mild 2 pixel erosion and computed the median depth value (mask median) as a robust estimate of fish depth. Depth sampling. Within the fish mask we applied a mild 2 pixel erosion and computed the median depth value (mask median) as a robust estimate of fish depth. Velocity calculation. Pixel displacements and depth were converted into world coordinate displacements, from which two and three dimensional swimming speeds (mm s⁻1) were calculated, with optional smoothing. Velocity calculation. Pixel displacements and depth were converted into world coordinate displacements, from which two and three dimensional swimming speeds (mm s⁻1) were calculated, with optional smoothing. Anomaly detection. Frames with three-dimensional speed above 1000 mm s⁻1 were flagged as anomalies and marked for inspection. Anomaly detection. Frames with three-dimensional speed above 1000 mm s⁻1 were flagged as anomalies and marked for inspection. Visualisation and export. The pipeline produced annotated videos, speed curves, smoothed speed curves, three dimensional trajectory plots, frame by frame CSV files and text summary reports. Visualisation and export. The pipeline produced annotated videos, speed curves, smoothed speed curves, three dimensional trajectory plots, frame by frame CSV files and text summary reports. Assumptions for trajectory reconstruction Within this pipeline we reconstructed three dimensional swimming trajectories of P. swirei from monocular video under the following idealised assumptions about camera geometry, seafloor structure and fish motion: The camera remained fixed during filming. The camera remained fixed during filming. P. swirei occupied a relatively small portion of each frame and its outline varied only modestly during swimming, so the displacement of the fish could be approximated by tracking the centre of its segmentation mask. P. swirei occupied a relatively small portion of each frame and its outline varied only modestly during swimming, so the displacement of the fish could be approximated by tracking the centre of its segmentation mask. The frame rate was high relative to the swimming speed, so motion between adjacent frames could be approximated as linear. The frame rate was high relative to the swimming speed, so motion between adjacent frames could be approximated as linear. The centre of the fish in each frame was defined as the average of the minimum and maximum pixel coordinates in the horizontal and vertical directions within the fish mask. The centre of the fish in each frame was defined as the average of the minimum and maximum pixel coordinates in the horizontal and vertical directions within the fish mask. Only a single individual of P. swirei appeared in each analysed frame. Only a single individual of P. swirei appeared in each analysed frame. The rods and tape visible on the lander were approximately parallel to the lower edge of the video, which we used as a visual reference for horizontal orientation. The rods and tape visible on the lander were approximately parallel to the lower edge of the video, which we used as a visual reference for horizontal orientation. Optical distortion of the camera lens was negligible and optical properties remained constant during the recording. Optical distortion of the camera lens was negligible and optical properties remained constant during the recording. The seafloor within the field of view was approximated as a planar surface. The seafloor within the field of view was approximated as a planar surface. The camera lens was treated as a point source, consistent with a pinhole camera model. The camera lens was treated as a point source, consistent with a pinhole camera model. Segmentation and tracking To detect and track P. swirei in each video frame we used a segmentation network enhanced with a convolutional block attention module (CBAM)48. CBAM adds both channel and spatial attention to the backbone network, which improves feature representation in low contrast underwater imagery and helps delineate object edges and textures more clearly. For every frame the model generated a binary mask of the fish, a bounding box and the corresponding centre coordinates. These outputs provided the inputs for subsequent depth sampling and velocity calculation in the analysis pipeline. The segmentation network used to detect and segment P. swirei was trained on annotated underwater images with a CBAM attention module. The full dataset contained 868 manually annotated instances of P. swirei. These were split in an 8 to 2 ratio, with 694 samples used for training and 174 samples used for validation. A separate clipped video was used as the test set for video segmentation. Training details and performance curves are provided in Supplementary Figure S1. Depth estimation The relative camera distance (depth value) for each frame was estimated using the Depth Anything model with a Vision Transformer based encoder (ViT B; weights depth_anything_vitb14.pth). The model produced a pixel level depth map Z(x, y) for every frame. Within the fish segmentation mask we applied a mild 2 pixel erosion to remove edge pixels, then computed the median depth value (mask median). This median depth was taken as the representative depth of P. swirei in that frame, which improved robustness and reduced the influence of edge noise. Geometric relationships from pixel to camera coordinates We used a pinhole imaging model to convert pixel coordinates (x, y) in the image plane to camera coordinates (X, Y, Z). Under this model, the mapping is: where f is the effective focal length in pixels, (cx, cy) are the principal point coordinates (approximated by the image centre when calibration is not available), and Z is the depth in millimetres. The three-dimensional displacement between successive frames and in camera coordinates is: Velocity calculation (2D and 3D) The video frame rate, denoted FPS, defines the time interval between consecutive frames as: From the camera log, the spatial calibration was pixels per millimetre. The corresponding millimetre per pixel scale is: The two-dimensional displacement in the image plane between frames is given by the pixel shifts. and . The corresponding two-dimensional swimming speed is: The three-dimensional swimming speed in camera coordinates, based on the projected displacements , and , is: Smoothing and statistics To reduce frame to frame noise in the velocity estimates, we applied a simple moving average filter with an odd window length W (here ), using reflection padding at the boundaries to limit edge artefacts. The smoothed velocity time series v ~ (t) is: For each trajectory and across the population, we then computed descriptive statistics of the speed, including mean, median, minimum, maximum and standard deviation, together with the number of sampled frames. Where relevant, we also recorded the number of valid and invalid frames to quantify the effect of frame exclusion criteria. Anomaly detection To reduce the influence of tracking and depth effects, we applied an anomaly detection step to the three-dimensional speed series. Frames with estimated three-dimensional speed greater than 1000 mm s⁻1 were classified as anomalies. This value was chosen based on preliminary runs and visual inspection. Speeds above 1000 mm s⁻1 always coincided with single frame jumps, partial masks or obvious depth glitches, and did not match the continuous motion seen in the annotated videos. In contrast, values below this threshold formed smooth trajectories that followed the fish in a biologically plausible way. In the annotated videos anomalous frames were highlighted with a red bounding box around the fish, and the pipeline exported still images of these frames together with a text report. These outputs were used for manual inspection and to exclude clearly spurious points from the summary statistics and ecological analysis. An example frame flagged by the anomaly detector with full contextual information (speed, pixel displacement and depth change) is shown in Supplementary Fig. S3. Visualization and rendering For each analysed sequence we generated a standard set of visual outputs. At the frame level, we overlaid the segmentation mask, bounding box and centre point of P. swirei together with the current depth and three-dimensional speed values. Raw 3D trajectories used to validate the reconstruction pipeline are shown in Supplementary Fig. S4. Frames flagged as anomalous by the threshold rule were additionally highlighted in red. We also produced paired views of the original RGB image and the corresponding depth map rendered with a JET colour scale. At the time series level the pipeline exported plots of three-dimensional speed against frame number, including both the raw and smoothed curves with embedded summary statistics. Finally, reconstructed trajectories were rendered as three dimensional paths in camera coordinates, together with orthogonal projections in the XY, XZ and YZ planes. Odor plume and fish response model We estimated the distance at which P. swirei first detected the bait odor with a two-phase framework that couples near bottom advection diffusion of the odor plume to a biased random walk model of fish movement. The total duration between bait deployment and the end of each analysed sequence, denoted T, was determined from the video timestamps and partitioned into an odor plume propagation phase of duration tb​ and a directed response phase of duration tf ​, such that: During the plume propagation phase the maximum down current distance of the odor front was written as: Here, u is the mean near bottom current velocity, k is a dimensionless coefficient that sets the effective edge of the plume, and is the effective diffusion coefficient that combines molecular diffusion Dm and turbulent diffusion Dt. We used fixed physical parameters based on hadal near bottom conditions: , , and . Once the fish first encountered the plume at distance R, its approach to the bait was treated as a biased random walk during the response phase . In this framework, plume encounter was not determined from a behavioral change in the video but was estimated mathematically as the point where plume propagation distance equals the modeled response trajectory of the fish. The net displacement towards the bait was written as: The effective chemotactic speed is , where v is the mean swimming speed and x is an orientation coefficient between 0 and 1 that measures how often the fish moves in the direction of the bait. The fish diffusion term is , where γ is the turning frequency and is the number of spatial dimensions. Values of , γ and x were measured from each reconstructed trajectory and ranged from to for v, to for γ and to for x. The core assumption of the model is that the distance travelled by the plume to reach the fish is equal to the distance the fish then travels back to the bait source. This gives the intersection condition: Together with the relation Eq. (8), Eq. (11) defines a unique solution for the plume propagation time tb: From this we obtained an interval of possible initial positions at bait deployment. The numerical values of R, Lrandom and the initial position interval are reported in the Results (Sect. 2.3). Accordingly, the estimated odor detection distance should be interpreted as an approximate, model-derived value rather than a directly observed quantity, and its variance cannot be robustly quantified with the current dataset. Conclusions This study establishes a framework for observing the behavior of hadal animals directly in their natural environment and demonstrates that such observations can yield quantitative biological insight even at the deepest ocean depths. By reconstructing three-dimensional motion from single camera video, we show that it is possible to analyze movement patterns, spatial use and responses to environmental cues without disturbing the animals or altering the conditions in which they live. The ability to extract reliable trajectories from deep sea video opens opportunities that extend well beyond the Mariana snailfish. Similar workflows can be applied to investigate diel activity, interactions with prey and scavengers, responses to currents and seafloor features, and variation in behavior across trenches and environmental gradients. As imaging systems, autonomous platforms and machine learning tools continue to improve, it should become possible to build long term behavioral datasets for hadal organisms and integrate these with genetic and environmental information. Such an effort would allow researchers to examine how animals function under extreme pressure, how energy moves through hadal food webs and how deep ocean communities may respond to future environmental change. Supplementary Information Supplementary Information 1. Supplementary Information 2. Acknowledgements This work was supported by the project 2025YFE0218900. We are grateful to R/V Tansuo Yihao for their contributions in lander deployment. Author contributions A.S.B. analyzed the data and wrote the first draft of the manuscript. Y.C. performed the video analysis, frame-based calculations, and prepared the figures. J.Z., J.L., L.H., and A.S.B. reviewed the manuscript. Y.W. conceived the study. Funding This study was supported by National Natural Science Foundation of China (42376149) and Shenzhen Key Laboratory of Advanced Technology for Marine Ecology (ZDSYS20230626091459009). Data availability The computer-vision analysis pipeline used for fish detection, monocular depth estimation, trajectory reconstruction, and swimming-speed calculation is openly available on GitHub at: https://github.com/chenyizi65/FishSpeed. The repository contains the scripts and configuration files required to reproduce the analyses. Data generated during this study are derived from deep-sea lander video recordings and processed outputs produced using this pipeline. Representative data are included in the manuscript and Supplementary Information. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Yizi Chen and Alexandru S. Barcan. References Associated Data Supplementary Materials Supplementary Information 1. Supplementary Information 2. Data Availability Statement The computer-vision analysis pipeline used for fish detection, monocular depth estimation, trajectory reconstruction, and swimming-speed calculation is openly available on GitHub at: https://github.com/chenyizi65/FishSpeed. The repository contains the scripts and configuration files required to reproduce the analyses. Data generated during this study are derived from deep-sea lander video recordings and processed outputs produced using this pipeline. Representative data are included in the manuscript and Supplementary Information.