Visual uncertainty and task demands shape active sensing strategies in mice

Abstract

In natural environments, animals actively sample visual information to guide behavior. Sensory feedback is dynamic and often requires active movements, whether saccading across the lines of this page or walking through a park. From high-acuity vision in hawks to low-acuity mice, many animals actively navigate to seek information, which can be called infotaxis. Although mice have relatively low-acuity vision, they still rely on sight for critical behaviors including navigation and prey capture. Yet, how sensitive they are to visual information and performing infotaxis has not been established. Here, we develop a virtual reality object discrimination task to investigate visual decision-making under naturalistic conditions. We show that mice perform infotaxis by actively seeking out informative views to guide their choices. Stimulus manipulations confirm that this strategy is modulated by the amount of available visual information. These results reveal that mice use principled active strategies to resolve visual uncertainty, highlighting a key role for information-seeking in natural vision.

Introduction

Traditional vision studies have largely assumed a passive observer (1–6), yet animals in natural environments actively gather visual information to guide behavior (7, 8). This process, called active sensing, involves the use of movement to selectively sample sensory input in a goal-directed manner (9–11). From a computational normative perspective, active sensing forms a closed loop between an observer, which infers task-relevant variables from incoming sensory data, and a planner, which selects actions that yield informative observations (12). This framework connects sensory processing and motor behavior through continuous, goal-directed information gathering (13–15).

The implementation of active sensing varies dramatically across species depending on sensory anatomy, leading to distinct sampling strategies. Species with a high-resolution portion of the retina, such as a fovea or area centralis, benefit from rapidly shifting that high-acuity region via eye movements (16). In primates, this is characterized by rhythmic sampling of the visual scene where saccadic eye movements and stereotyped fixation durations aligns the fovea with task-relevant visual features (9, 17–22). By contrast, species such as mice with more uniform retinal resolution (23, 24) rely more heavily on head and full body movements to leverage cues such as motion and position parallax, while their eye movements occur mostly in tandem with head movements to compensate for reorientation (25–27). Despite these anatomical differences, mice still depend on vision for a range of ecologically relevant behaviors including threat detection (28, 29), foraging (30, 31), spatial navigation (32, 33), and depth perception (27, 34). However, no paradigm has yet been developed to systematically measure specific active visual infotaxis strategies in this species (11).

To address this gap, we developed a free-range psychophysics paradigm for laboratory mice to test the hypothesis that mice engage in active visual sensing based on task demands. We designed a virtual reality-based object discrimination task that used real-time pose estimation to control visual stimuli, enabling dynamic manipulation of visual occlusion via a wall-like aperture that introduced depth parallax. Specifically, changes in viewing angle exposed different parts of the object from behind occluders: by moving closer to the screen, mice could reveal more of the occluded object. Although head-fixed virtual reality approaches enable precise visual control at the cost of natural movements (35–37), our freely moving paradigm allowed mice to actively control their viewing angle through body and head movements, enabling them to actively sample visual information while we maintained complete knowledge of the visual stimuli. We found that mice modulated their behavior based on the amount of available information, consistent with an infotaxis strategy that seeks to maximize information gain (7, 38–40). Trial-by-trial decisions could be decoded from head-body movements and locomotor speed, with decoding accuracy varying based on object visibility. This indicated that mice adapted their sensing strategies in response to visual uncertainty. These findings demonstrate that mice perform active visual sensing and flexibly deploy sensorimotor strategies to resolve object ambiguity under variable visibility conditions.

Results

An active sensing task for mice.

To investigate whether mice engage in active sensing during visual decision-making, we developed a freely moving virtual reality (VR) system that provides real-time, closed-loop visual feedback from the point of view of the animal, based on its position and posture (Figure 1A, STAR Methods). We used an off-axis projection of the scene (Figure 1B) so that visual scenes were rendered dynamically from the instantaneous viewpoint of the animal and the correct motion parallax was preserved. Nearer virtual objects shifted more with head movements than distant ones, producing a realistic three-dimensional visual experience. To generate the scene from the animal’s point of view, we integrated DeepLabCut-Live! for real-time pose estimation (41) (Figure S1A:C) with Unity for instantaneous viewpoint rendering (Figure S1D), managed via a custom Python-based bridge (Figure 1C). This ensured low-latency, closed-loop feedback (mean ± s.e.m. per lab: Tolias lab: 80.62 ± 2.40 ms, n = 5; MLAI: 56.10 ± 0.83 ms, n = 6; Niell lab: 59.70 ± 0.40 ms, n = 5; Figure 1D, Figure S1E:G). We also controlled for acoustic stability of the experimental environment and showed that the system does not emit task-aligned noise that could help the mice in any way (Figure S1H). In brief, this system enabled naturalistic interaction with visual stimuli while maintaining precise experimental control, allowing us to design tasks that probe visual information-gathering behavior. We implemented identical setups and protocols across three laboratories, establishing a reproducible framework for studying active vision.

Figure 1. A closed-loop virtual-reality active sensing task for mice.

A: Schematic of the virtual reality arena. The mouse moves freely around while a virtual world renders dynamically on the front screen. Lick ports are situated at the front sides. A high-speed overhead camera acquires images at 100Hz. B: On- vs. Off-axis projection: Above, On-axis projection: camera viewing angle moves out of projection plane. When the mouse moves right, the camera frustum remains symmetric relative to the center, leading to inaccurate motion parallax. Below, Off-axis projection: camera viewing angle is fixed to the projection plane. When the mouse moves right, the projection matrix adjusts to its viewpoint, accurately aligning virtual content to its perspective. C: Virtual reality software architecture. DeepLabCut-Live captures images for real-time pose estimation. Head x/y positions and running direction are extracted via our custom Python toolbox, vr4mice, and sent to Unity for rendering. vr4mice also controls water reward delivery via a Teensy 4.0 for correct reports. D: Total camera-to-render latency across labs, summing DeepLabCut-Live and Unity latencies (Figure S1E:G). Mean ± s.e.m.: 80.62 ± 2.40 ms (Tolias; n=5), 59.70 ± 0.40 ms (Niell; n=5) and 56.10 ± 0.83 ms (MLAI; n=6). E: Arena top view overlaid with the trial-start area (blue) and report areas (orange: right; purple: left). Areas not visible to the mouse. Main screen at the front. F: Trial structure. A trial starts when the mouse enters the initiation area, aligns its running direction with the screen, and meets a velocity threshold. The mouse reports the object of interest location by entering the corresponding side reward port. G: Example trial trajectories from a trained mouse during discrimination (session: Oribi, 2024–08–14). Blue/Red dots: start/end positions. See also Figure S1.

To initiate a trial, the mouse had to enter a trial initiation zone (non-visible to the mouse) in the arena while decreasing its velocity and facing the front screen for 250 ms (blue rectangle in Figure 1E). This caused two objects to appear on the screen in front of the mouse. It could then report the location of the target object by going to the right or left lick port (purple and orange shaded rectangles in Figure 1E). If the mouse reported correctly by going to the lick port on the same side as the target object, it received a water reward (3–5 µL), and the objects disappeared until the mouse initiated another trial (inter-trial interval). If the mouse reported incorrectly, no water reward was delivered and the inter-trial interval was triggered directly (Figure 1F). Across the three laboratories, on the majority of trials mice displayed trajectories that formed a J-shape as they moved toward the lick ports (Figure 1G).

Mice were trained in a two-step regimen (Figure 2A, Figure S2A): first an object detection task, followed by an object discrimination task. In the detection phase, mice learned to report the correct location of an object of interest (a white teardrop). The object is stationary in the virtual coordinate system, but the rendered viewpoint is coupled to the mouse’s movements so that the available visual information at any given point in space depends on the mouse’s position. To facilitate learning trial initiation, the detection phase was split into two subtasks. First, the mice had to enter the start zone and remain for 100 ms to initiate a trial. Once they reached sufficient performance (above a 70% reward rate on more than 125 trials within an hour, see STAR Methods), they were switched to the standard trial-initiation criteria used for the rest of the paradigm: entering the start zone for a longer duration of 250 ms. After reaching criterion on this as well, they graduated to the discrimination task, where the target object and a distractor object (black teardrop) were presented simultaneously, and mice had to report the target’s location.

Figure 2. Mice can learn to discriminate objects based on luminance in a motion-parallax correct virtual reality world.

A: Mice first learned to detect a single white teardrop (detection phase). Next, a black teardrop distractor was added (discrimination phase); mice still reported the white teardrop’s position. B: Success rate across training phases. Green: detection phase, showing first/last sessions without velocity threshold (First/Middle), and the last session with velocity threshold (Last). Pink: first, middle (if applicable), and last sessions of the discrimination phase (with distractor). One-way ANOVA, significant training stage effect: F(1, 53) = 8.13, p = 0.006. Tukey HSD tests, significant pairwise differences, ∗∗∗: p = 0.001, ∗∗: p = 0.002, ∗: p = 0.025 (n=10), Figure S2B. C: Averaged training days (Left) and corresponding total number of trials (Right) until mice were considered experts at the discrimination task. Mean ± s.e.m.: 13.4 ± 1.76 days, 1644.0 ± 143.6 trials (n=10). D: Mean J-shaped trials rate across training. One-way ANOVA, significant training stage effect: F(1, 53) = 29.53, p < 0.001. ∗: Tukey HSD tests, significant pairwise differences between first detection session and others, p < 0.001 (n=10), Figure S2D. See also Figure S2.

Throughout training, mice showed an increase in performance, with a significant improvement occurring between early and late training sessions for the detection and discrimination phases (one-way ANOVA significant effect of training stage, F(1, 53) = 8.13, p = 0.006; Tukey HSD, First vs. Last detection p = 0.025, First vs. Last discrimination p = 0.002, First detection vs. Last discrimination p = 0.001; Figure 2B, Figure S2B). Mice reached expert performance (same criterion as for detection phase: 70% success rate on at least 125 trials within 60 minutes, see STAR Methods; number of trials per session significantly higher than the first training session for all subsequent sessions; Tukey HSD p < 0.001; Figure S2C) in 13.4±1.76 days for a total of 1644±143.6 trials on average (mean ± s.e.m.; Figure 2C). Following the first detection session, J-shaped trials became significantly more frequent (Tukey HSD, p < 0.0001; Figure 2D, Figure S2D), while trial tortuosity decreased significantly (Figure S2E; Tukey HSD p < 0.0001, see STAR Methods). Simultaneously, reporting time dropped significantly (Tukey HSD, First vs. Middle detection p = 0.009, overall p < 0.025; Figure S2F), while a significant choice bias emerged transiently upon presentation of the distractor (one-sample t-test, First discrimination: t = 3.03, p = 0.014; Middle discrimination: t = 5.20, p = 0.007, all others p > 0.25; Figure S2G). Overall, these results confirm that mice reliably learned a consistent set of actions consisting of J-shaped trajectories to solve the task.

Active sensing under partial visibility of objects.

To investigate whether mice could perform visual active sensing, we introduced a task where partial occlusion of the objects occurred by presenting them behind wall-like occluders of randomly-sampled size (99% versus 10% of objects visible when viewing from the center of the trial initiation area, Figure 3A; equal frequency aperture conditions: paired t-test, t(42) = 0.49, p = 0.63, Figure S3A; Video S1). We hypothesized that this manipulation would increase task demands and require mice to modulate their behavioral strategy accordingly. Specifically, if mice perform active sensing, they should adjust their information-seeking behavior in response to reduced object visibility, upon the narrow-aperture condition, as less visual information would initially be accessible through the aperture.

Figure 3. Mice stay more central, move slower and orient themselves more toward the objects as they approach the screen when faced with obstructed objects.

A: Extent of the visual aperture during the dual-occlusion task: wide (99% of the objects visible) vs. narrow (10% of the objects visible). B: Success rate per aperture. Mean ± s.e.m.: 0.76 ± 0.01 (narrow) and 0.86 ± 0.01 (wide). Paired t-test: t(42) = 6.98, p = 1.55 ×10⁻⁸. C: Influence of previous trial outcome on current choice. Above: Side choice bias per aperture. 0 = no bias, ±1 = full Left/Right. Mean ± s.e.m.: 0.158 ± 0.042 (narrow) and 0.052 ± 0.035 (wide). One-sample t-test vs. no bias across mice: narrow, t = 2.09, p = 0.066; wide: t = 1.07, p = 0.311. Below: Decision stickiness per aperture. 0 = no stickiness to the previous decision, +1 = repeat, −1 = switch. Mean ± s.e.m.: 0.025 ± 0.021 (narrow) and 0.033 ± 0.018 (wide). One-sample t-test vs. no stickiness across mice: narrow, t = 0.01, p = 0.995, wide, t = 1.08, p = 0.307. D: 3 × 3 initial position distribution across trials (n=10 mice, s=43 sessions). See Figure S3E for choice-specific distributions. E: Example trial trajectories for wide (left) and narrow (right) apertures (session: Pheasant, 2024–08–15). Mean trajectory ± s.e.m. (shaded) per aperture and choice for F: individual example mouse (Pheasant, s=5 sessions; Figure S3F), and G: mean across mice (n=10 mice, s=43 sessions). Two-way repeated-measures ANOVA, significant effects of aperture: F(1, 42) = 68.65, spatial bin: F(40, 1680) = 89.21, and interaction: F(40, 1680) = 6.41; all p < 0.0001. ∗: FDR-corrected paired t-test with multiple comparisons across y bins (0–18 cm in the arena), significant aperture effect (p < 0.05) from 0 cm onward; Figure S3G.Trajectories shown from arena midpoint (y-axis) to screen; dashed line indicates x-axis center. H: Mean velocity ± s.e.m. per aperture (n=43). Two-way repeated-measures ANOVA, significant effects of aperture: F(1, 42) = 58.36, trial progression: F(99, 4158) = 48.18, and interaction: F(99, 4158) = 35.61; all p < 0.0001. ∗: FDR-corrected paired t-test with multiple comparisons across trial progression, significant aperture effect (p < 0.05) on the 40–80% portion of the trial, Figure S3H. I: Depending on the mouse’s forward-movement toward the screen, we can approximate a trajectory using a constant ${\text{L}}_p$ norm curve, parameterized by $p$ and the horizontal and vertical length scales a and b. Larger $p$ indicates sharper curvature, reflecting further forward movement before committing to a decision under narrow occlusion. J: Optimal $p$ parameter from single-trial ${\text{L}}_p$ curve fits. Mean ± s.e.m.: 16.68 ± 0.58 (narrow) and 14.53 ± 0.49 (wide). ∗∗∗: paired t-test, t(42) = −7.12, p = 9.78 ×10⁻⁹. For B, C and J: Black: mean ± s.e.m.; colored: per-mouse; grey: per-session. See also Figure S3 and Video S1.

Mice maintained high discrimination accuracy in both occluded and unoccluded trials (mean narrow = 0.76 ± 0.01; mean wide = 0.86 ± 0.01; Figure 3B, Figure S3B). The significant decrease in accuracy during occlusion (paired t-test: t(42) = 6.98, p = 1.55 ×10⁻⁸) effectively highlights the increased sensory complexity posed by hidden objects while confirming the mice’s ability to solve more demanding conditions. They localized the object positions without a population-level side bias (one-sample t-test vs. no bias: narrow, t = 2.09, p=0.066; wide: t = 1.07, p = 0.311; mean narrow: 0.129 ± 0.062, mean wide = 0.069 ± 0.064), or decision stickiness (one-sample t-test vs. no stickiness to the previous choice; narrow: t = 0.01, p = 0.995, wide: t = 1.08, p = 0.307) at both aperture levels, even if subtle individual biases existed (Figure 3C; Figure S3C for results per laboratory). High rates of J-shaped trials also indicate strong task engagement (mean: 0.972 ± 0.0044; Figure S3D). Furthermore, we found no difference in the trial initial position between choice sides (paired t-test: FDR-corrected p-value, p > 0.05 for all spatial bins; Figure 3D, Figure S3E). This suggests that mice naturally move toward a specific position before task initiation, resulting in a consistent starting view of the objects regardless of the subsequent decision. This consistent engagement and lack of systematic bias – even under occlusion – sets the stage for examining underlying information-seeking strategies.

A behavioral strategy that would enable mice to perform the task accurately when faced with occlusion is to move closer to the screen while remaining near the arena center. This would result in a wider viewing angle through the aperture, allowing the mice to observe more surface area of the objects of interest. We observed such a pattern across trials (Figure 3E) and in the per-mouse averaged trajectories (Figure 3F, Figure S3F). As they moved toward the front screen, all mice stayed in the central portion of the arena more in the narrow than in the wide aperture condition. Overall, we observed a significant difference between averaged trajectories per aperture condition, consistent with our initial hypothesis (two-way repeated-measures ANOVA significant effect of aperture: F(1, 42) = 68.65, spatial bin: F(40, 1680) = 89.21, and interaction: F(40, 1680) = 6.41; all p < 0.0001; FDR-corrected paired t-test with multiple comparison across y bins: p < 0.05; Figure 3G, Figure S3G).

In addition, if movements were independent of the information available, both conditions should exhibit identical velocity profiles, and if mice were unsure about their decision, this should be reflected in their velocity. Analysis of the animals’ velocity revealed a significantly slower acceleration between 40% and 80% of trial progression in the narrow aperture condition compared to the wide aperture trials (two-way repeated-measures ANOVA significant effects of aperture: F(1, 42) = 58.36, time progression: F(99, 4158) = 48.18, and interaction: F(99, 4158) = 35.61; all p < 0.0001; FDR-corrected paired t-test with multiple comparison across y bins: p < 0.05; Figure 3H, Figure S3H). When examining the velocity components independently, the y-velocity (forward movements, toward the screen; Figure S3I) was lower throughout the trial under occlusion, while the x-velocity (lateral movements, toward the reward ports; Figure S3J) showed a delayed peak in the narrow condition. This indicates that the increased sensory demand of the narrow aperture led to a more cautious velocity profile characterized by a slower approach to the screen and a delayed commitment to a choice side, likely to maximize the time available for information gathering.

Beyond velocity, the spatial orientation of the mice also diverged according to the available information. The running direction differed significantly between conditions from trial onset, except between 70% and 80%, which aligns with the observed velocity decrease in the narrow condition (two-way repeated-measures ANOVA, significant effects of aperture: F(1, 42) = 19.06, time progression: F(99, 4158) = 1344.43, and interaction: F(99, 4158) = 33.54; all p < 0.0001; FDR-corrected paired t-test with multiple comparisons: p < 0.05; Figure S3K-L). This suggests that mice adapted their behavioral trajectory from the very beginning of the trial based on the presented level of occlusion. Moreover, the angle of the head relative to the body showed significant differences from 60% to 90% of the trial (two-way repeated-measures ANOVA, significant main effects of aperture: F(1, 42) = 30.96, time progression: F(99, 4158) = 92.55, and interaction: F(99, 4158) = 5.83; all p < 0.0001; FDR-corrected paired t-test with multiple comparisons: p < 0.05; Figure S3M-N). This postural shift in the later stage is consistent with the sudden turn towards the reward port observed in the narrow condition’s average trajectory, suggesting a rapid commitment once sufficient information has been integrated.

Taken together, these results suggest that when faced with a high degree of occlusion (narrower aperture), mice moved closer toward the screen for longer in the trial before making a decision. One possible reason for this is that mice move to a position within the arena where they increase their visual angle through the aperture, allowing them to better view the objects of interest to then correctly classify their location. As such, these trials can be thought of as having two distinct phases: a search phase where mice seek out object information that is hidden by the occlusion, followed by a decision phase where – once enough information has been integrated (i.e., once they are confident enough) – they commit to going to one of the reward ports. In the wide-aperture trials, on the other hand, mice can readily obtain this information from further back in the arena and can commit to a decision without entering a search phase, or at least with only a shortened one.

To better quantify this, we fitted the individual experimental trial movement trajectories to ${\text{L}}_p$ curves by finding the optimal $p$ for each trial (STAR Methods). The ${\text{L}}_p$ family of curves provides a flexible way to capture trajectory curvature, where larger $p$ values correspond to more sharply curved (i.e., more J-shaped) paths, consistent with the intuition that mice exhibit more pronounced J-shaped trajectories in narrow aperture trials (Figure 3I). We observed a significant difference in the average optimal p between aperture conditions (mean narrow = 16.68 ± 0.58, mean wide = 14.53 ± 0.49; paired t-test across sessions t(42) = −7.12, p = 9.77 ×10⁻⁹; Figure 3J, Figure S3O), indicating sharper turns when objects are occluded. Similarly, the significantly higher tortuosity observed in narrow aperture trials compared to wide ones (mean narrow = 1.56 ± 0.03, mean wide = 1.33 ± 0.01, paired t-test: t(42) = − 6.85, p = 2.41 ×10⁻⁸; Figure S3P) indicated that mice shifted to more complex search paths when sensory information is constrained. Analysis of the most tortuous forward paths (top 90^th percentile) revealed that these were predominantly narrow-aperture trials (83.6%; Figure S3Q). These trajectories sometimes showed “change of mind” patterns, characterized by an initial movement toward one target followed by a correcting redirection toward the other. Such behavior indicated that the trajectories were not ballistic or pre-planned. Instead, they suggested that mice actively adjusted their behavior in reaction to late-arriving evidence, utilizing incoming information to update an internal decision variable and self-correct mid-trial (42, 43).

We also trained a batch of mice (n=3) on the inverse contrast experiment (black target, white distractor) and observed similar behavioral changes (Success rate: narrow: 0.74 ± 0.02, wide: 0.90 ± 0.02; t(14) = 4.35, p = 6.67×10⁻⁴; Trajectory: two-way repeated-measures ANOVA, significant effects for aperture: F(1, 14) = 11.96, p = 0.0038, spatial bin: F(14, 196) = 221.97, p < 0.0001 and interaction: F(14, 196) = 5.80, p < 0.0001; FDR-corrected paired t-test with multiple comparisons across y bins, significant aperture effect from 2.5 cm of the initiation area top onward; Combined velocity (m/s): two-way repeated-measures ANOVA, significant effect of aperture: F(1, 14) = 63.07, time bin: F(99, 1386) = 46.39, and interaction: F(99, 1386) = 22.22; all p < 0.0001; paired t-test with multiple comparisons across the trial progression, significant aperture effect at 40–80% trial; Decision points distance to the screen (cm): narrow: 14.64 ± 0.67, wide: 18.29 ± 0.52; t(14) = 3.76, p = 0.021; STAR Methods; Figure S3R). This suggests that the animals’ navigation and sensing strategies are driven by object identity and task rules rather than a luminance-based detection of the white object or confusion of the black object with the arena walls.

A dynamic choice variable predicts infotaxis behavior.

We hypothesized that aperture width should shift the position at which mice commit to a choice: in narrow-aperture trials, decisions should occur closer to the screen than in wide-aperture trials. To resolve the time course of infotaxis, we trained a time-resolved multi-model framework consisting of ten independent logistic regression models – one per 10% window of the trial – to predict the animal’s upcoming choice (left versus right reward port). One model was trained per time segment and session in a leave-one-session-out procedure (n=43; Figure 4A; STAR Methods). This allowed us to estimate when behavior first became predictive of the final decision. To determine which features best capture the decision process, we compared ten candidate feature sets using the Bayesian Information Criterion (BIC; STAR Methods). The best-performing model included only kinematic and orientation features related to lateral displacement (Figure S4A; Figure S4C:F for regressor time courses). Thus, lateral movements provide the most informative behavioral signature of choice formation. This is consistent with the task geometry, where the x-axis defines the left–right decision axis in the arena. Analysis of regression weights along the time segments revealed a three-stage evolution in the relative importance of these features (Figure 4B, Figure 4G): in the beginning of the trial (0–30%), predictive power is negligible; during the mid-trial deliberating phase (40–80%), x-velocity and orientation emerge as the primary predictors; finally, as the mice approach the choice (80–100%), x-position becomes the dominant feature.

Figure 4. Behavioral features reliably define a dynamic choice variable revealing that, under occlusion, mice rapidly shift decision points to more informative viewpoints, correlating with better performance.

A: Logistic regression setup. Four kinematic features (lateral position, velocity, head-body angle, and running direction) selected via BIC analysis to predict the choice side (Figure S4A). Ten independent models trained per session on non-overlapping 10% trial segments using leave-one-session-out cross-validation. B: Mean logistic regression weights (logits) ± s.e.m. across sessions (n=43 sessions) and trial segments (n=10 segments; segment-specific weights: Figure S4G). Dynamic choice variable (P(left)) for C: example trials, and D: mean ± s.e.m across trials and models per choice and aperture (n=43 sessions). Grey lines: decision threshold at a 20% confidence bound (STAR Methods). E: Model prediction accuracy per aperture. ∗: FDR-corrected paired t-test with multiple comparisons across trial progression, significant aperture effect (p < 0.05) on 30–90% of trial progression, Figure S4H. F: Example trial trajectories colored by instantaneous dynamic choice variable; pink dots indicate decision points (20% confidence bound). Maps display object visibility across arena positions per aperture (STAR Methods). G: Single trajectory example, with stimuli at onset, decision, and reward. Decision at 36% (wide) and 44% (narrow) of trial. H: Decision point distance to screen (cm) per aperture and mouse. Means ± s.e.m.: narrow: 13.79 ± 0.44 cm, wide:17.54 ± 0.42 cm (difference: 3.75 cm). ∗∗∗: paired t-test, t(42) = 6.03, p = 3.56 ×10⁻⁷. Black: mean ± s.e.m.; colored: per-mouse; grey: per-session. I: y-velocity vs. information content aligned to decision (from −1s) per aperture. Linear least-squares regression from −0.4s to decision (r: Pearson correlation, p: Wald test p-value, H₀: slope is zero): narrow: r = −0.68, p < 0.001; wide: r = 0.42, p = 0.007. Peak velocity: −0.22s (narrow) and −0.20s (wide). J: Linear least-squares regression between success rate and decision distance to screen (cm) per aperture: narrow: r = −0.709, p = 0.022; wide: r = −0.103, p = 0.777. Per-mouse average across all sessions, regardless of the reward-drop. K: Success rate (above) and decision distance (cm) (below) across the first occluded session (200 trials, n=10 mice) using a 15-trial moving average. Dotted: first 15-trial baseline; solid: mean across mice; thin: individual mice. See also Figure S4.

We used the time-specific probability estimates of these models as a dynamic choice variable (43, 44) along trial progression (Figure 4C-D). First, we observed that the prediction accuracy of the model throughout the trial per occlusion condition showed a significant shift between conditions from 30% to 90% (two-way repeated-measures ANOVA significant effects of aperture: F(1, 42) = 26.19, trial progression: F(49, 2058) = 794.25, and interaction: (F(49, 2058) = 9.63; all p < 0.0001; FDR-corrected p < 0.05; Figure 4E, Figure S4H). Thus, the model reached the same level of confidence significantly later in the trial under narrow apertures. This indicates that the behavioral features used as model regressors varied sufficiently between aperture conditions to drive this temporal predictability difference, suggesting distinct behavioral strategies.

To quantify the timeline of the choice process, we defined the decision point – i.e., the transition from information gathering to directed navigation – as the moment the model confidence reached a specific certainty threshold (Figure 4F-G). To determine the threshold that most accurately reflected the mice turning dynamics, we evaluated certainty levels ranging from 50% (random) to 90%. A threshold of 80% was found to best capture the onset of the animal’s lateral trajectory toward the reward port (Figure S4B; STAR Methods). Decision points defined as such occurred significantly closer to the screen in the narrow aperture condition compared to its wide counterpart (means ± s.e.m. narrow: 13.79 ± 0.44 cm and wide: 17.54 ± 0.42 cm, mean difference = 3.75. ∗∗∗: paired t-test, t(42) = 6.03, p = 3.56 ×10⁻⁷; Figure 4H, Figure S4I).

Overall, as task demands increased – with the narrow aperture increasing visual uncertainty and making the task more challenging to perform as evidenced by their drop in performance – mice delayed their decision, spending more of the trial getting closer to the front screen. This resulted in them moving to more informative viewpoints (i.e., more of the objects was visible; information gain between initial and decision point positions, paired t-test, t(42) = −10.97, p = 6.67 ×10⁻¹⁴; Figure S4J). We then analyzed how movement dynamics evolved as evidence accumulated during the search phase. Specifically, we examined the velocity toward the screen (y-axis) as a function of the information content aligned to the decision. In both conditions, velocity increased rapidly until approximately 0.4 seconds before the decision point. Their trajectories then diverged: in narrow-aperture trials, velocity peaks at 0.22 seconds before the decision before decreasing (r = −0.68, p < 0.001), while in wide-aperture trials, velocity peaks at 0.20 seconds before the decision but stays overall high (r = 0.42, p = 0.007; Figure 4I). Then both conditions showed an abrupt deceleration immediately preceding the decision, corresponding to the physical turn to the reward port. This demonstrates evidence-driven modulation of velocity, reflecting reactive behavioral adjustments to task difficulty and visual uncertainty and consistent with an extended infotaxis strategy when faced with object occlusion.

This infotaxis behavior was correlated with higher success rates under the narrow-aperture condition but not under the wide-aperture one: mice that moved closer to the screen to make a decision on average were also the most successful on the occluded discrimination task, and this, even though all of them already performed well (above 0.7 success rate threshold; linear least-squares regression across all sessions, independent of the reward drop between conditions: narrow: Pearson’s r = −0.709, Wald test, p = 0.022; wide: r = −0.103, p = 0.777; Figure 4J; paired t-test between the Closest and Furthest mice quintile – 20% – groups per aperture condition: narrow: t(41) = −3.16, p = 0.0029, wide: t(41) = 0.18, p = 0.86; Figure S4K). This suggests that moving the decision point closer to the screen is functionally relevant to task execution.

Interestingly, the immediate emergence of this strategy when mice were first exposed to occluders (in the very first dual occlusion test session; Figure 4K) suggests that they rapidly formed a robust internal model of the environment. Rather than gradually exploring or purely trial-and-error learning, they were able to apply this internal model efficiently, directly adjusting their trajectories and decision points to maximize information even under novel occlusion conditions. This indicates that mice can perform infotaxis not only reactively, but also proactively, leveraging an internal representation of the task structure to guide efficient exploration and decision-making from the very first exposure.

Infotaxis behavior correlates inversely with the amount of visible object area.

The dual-occlusion task demonstrated that mice modulate their trajectories and decision location when visual information is limited. However, this binary manipulation could not reveal whether their information-seeking strategy scales continuously with sensory uncertainty. We therefore developed a second occlusion task with five uniformly presented occlusion levels, ranging from fully visible objects (99%) to fully hidden objects (0%) when the animal starts the trial from the center of the initiation area (Figure 5A; balanced aperture condition appearance, paired t-test p > 0.05, Figure S5A; Video S2). Across all three laboratories, mice performed the task with high accuracy. We observed a significant performance drop as the aperture narrowed, reflecting the gradual increase in task demand as visual uncertainty increases (FDR-corrected paired t-tests: 99% and 72% object visible conditions significantly different from the other conditions; Figure 5B, Figure S5B), yet they could modulate their behavior on a trial-by-trial basis to reliably report the location of the correct object without a population-level significant bias across all aperture conditions (all aperture condition choice bias indices remained close to zero – i.e., no bias, single-condition t-test vs. no bias: all p > 0.05; Figure S5C).

Figure 5. The amount of information available to the mouse correlates inversely with infotaxis behavior.

A: Object visibility across five logarithmically distributed aperture levels in the multi-occlusion task, ranging from 99% to 0% of the objects visible (Above) and example trial trajectories colored by the instantaneous dynamic choice variable (pink: decision points). Information maps show object visibility across arena positions per aperture. B: Sucess rate per aperture condition and mouse. Repeated-measures ANOVA, significant aperture effect: F(4, 96) = 37.52, p < 0.0001. FDR-corrected paired t-test: p < 0.00025 for all but 10–99%, 0–35%, and 0–10% pairs; Figure S5B. Mean trajectory ± s.e.m. (shaded) per aperture and choice for C: individual example mouse (Nightingale, s=4 sessions; Figure S5D) D: mean across mice (n=9 mice, s=25 sessions). FDR-corrected paired t-test with multiple comparisons across y bins (0–19 cm in the arena), significant aperture and choice effects (p < 0.05) between narrower and wider aperture conditions; Figure S5E. Midpoint to screen shown; dashed line: x-axis center. E: Mean velocity ± s.e.m. per aperture across trial progression. FDR-corrected paired t-test with multiple comparisons across trial progression, significant aperture effect (p < 0.05) between narrower and wider aperture at 40–80% of the trial; Figure S5F. F: Optimal p parameter from single-trial ${\text{L}}_p$ curve fits. Repeated-measures ANOVA, significant aperture effect: F(4, 96) = 3.14, p = 0.018. FDR-corrected paired t-tests: non-significant for all but the 0–72% pair; Figure S5G. G: Mean regression logits ± s.e.m. across sessions (n=25 sessions) and trial segment (n=10 segments). H: Mean model prediction accuracy ± s.e.m. per aperture. Two-way repeated-measures ANOVA, significant effect of aperture: F(4, 96) = 10.92, trial progression: F(49, 1176) = 428.47, and interaction: F(196, 4704) = 4.35; all p < 0.0001. I: Decision distance to screen (cm; 20% threshold) per aperture. Repeated-measures ANOVA, significant aperture effect: F(4, 96) = 39.61, p < 0.0001. FDR-corrected paired t-tests, significant differences for all but 0–10% and 72–99% pairs; Figure S5R. J: y-velocity vs. information content aligned to decision (from −1s) per aperture. Linear least-squares regression from −0.4s to decision (r: Pearson correlation; p: p-value), and peak velocity timing. For B, F, and I: Black: mean ± s.e.m.; colored: per-mouse; grey: per-session. See also Figure S5 and Video S2.

To assess the behavioral correlates of information-seeking under varying occlusion levels, we analyzed the average trajectories per session (Figure 5C, Figure S5D). For all aperture widths, mice showed the same characteristic J-shaped trajectories as in the dual-occlusion task. The steepness of the curve varied systematically, forming a gradient from shallow (wider apertures) to steep (narrower apertures). Across all mice, we observed significant differences in the averaged trajectories between the narrower and wider aperture conditions (FDR-corrected paired t-test: p < 0.05 for pairs 0–99% and 0–72%, 28.3 cm away from the screen onward, and 0–35%, 10–99%, 10–72%, 10–35%, 16.2 cm away from the screen onward; Figure 5D, Figure S5E).

Consistent with increased deliberation, kinematic analyses revealed significantly lower velocities during the 40–80% portion of trials under narrow aperture conditions (FDR-corrected paired t-test: p < 0.05 for 0–99%, 10–99% and 35–99% pairs in 40–80% of the trial; Figure 5E, Figure S5F). Additionally, the optimal p parameter from fitting ${\text{L}}_p$ curves (repeated-measures ANOVA: significant aperture effect: F(4, 96) = 3.14, p = 0.018, FDR-corrected paired t-tests: non-significant for all but the 0–72% pair; Figure 5F, Figure S5G) and trial tortuosity (repeated-measures ANOVA: significant aperture effect: F(4, 96) = 11.85, p < 0.0001, FDR-corrected paired t-test: p < 0.009 for all but 72–99%, 10–35%, 0–35%, and 0–10% pairs; Figure S5H) were both more elevated in narrower aperture trials, reflecting more circuitous paths.

The logistic regression framework trained to decode choice from lateral displacement behavioral features throughout trial progression (Figure 5G; Figure S5I:P) replicated the patterns observed in the dual-occlusion task. In particular, the model reached high confidence later in the trial for narrower apertures (two-way repeated-measures ANOVA, significant effect of aperture: F(4, 96) = 10.92, trial progression: F(49, 1176) = 428.47, and interaction: F(196, 4704) = 4.35; all p < 0.0001; Figure 5H; model accuracy: Figure S5Q), decision points were significantly closer to the screen under higher occlusion (repeated-measures ANOVA, significant aperture effect: F(4, 96) = 39.61, p < 0.0001; FDR-corrected paired t-test: p < 0.004 for all but 0–10% and 72–99% pairs; Figure 5I, Figure S5R) and instantaneous velocity was modulated by the corresponding available information (Figure 5J). Collectively, we observed that mice can perform active visual sensing and are behaviorally sensitive to relatively small changes in visual occlusion. Critically, we observed the same trends on the black target experiment (Success rates: 0%: 0.741 ± 0.019, 10%: 0.825 ± 0.030, 35%: 0.906 ± 0.021, 72%: 0.961 ± 0.016, 99%: 0.944 ± 0.016; Trajectories: two-way repeated-measures ANOVA, significant effects for aperture: F(4, 36) = 4.25, p = 0.0064, and spatial bin: F(24, 216) = 135.89, p < 0.0001, interaction not significant: F(96, 864) = 1.03, p = 0.4042, Combined velocity (m/s): two-way repeated-measures ANOVA, significant effect of aperture: F(4, 36) = 8.02, p = 0.0001, time bin: F(99, 891) = 36.20, and interaction: F(396, 3564) = 4.93; all p < 0.0001; Decision distance to screen: 0%: 13.785 ± 0.688, 10%: 14.765 ± 0.612, 35%: 16.407 ± 0.664, 72%: 17.583 ± 0.924, 99%: 17.385 ± 0.685; Figure S5S).

Discussion

Here, we present evidence that mice engage in visual active sensing as defined by Yang et al. (13), who outlined three theoretical hallmarks: task-dependent movements, sensor- and actuator-dependent strategies, and history-dependent actions. Accordingly, mice robustly perform visual infotaxis. They move to acquire new views of the visual scene to overcome visual uncertainty and perform the task more accurately. Interestingly, this behavior is similar to the “peeking” observed in a predator avoidance task (45), where mice quickly move to acquire visual information while minimizing their own exposure. Furthermore, by analyzing the trajectories, we are able to quantitatively assess that both the decision point and the changes in behavior were dependent on the amount of visual information available, demonstrating sensor-dependent changes in behavioral strategies. The transition from more direct paths to more flexible ones as visual uncertainty increases – notably with the presence of “changes of mind” trials – aligns to active sensing strategies observed in other modalities, such as olfactory infotaxis, where animals switch from stereotyped tracking to sensory-driven search to resolve spatial ambiguity (46). Lastly, we show that these movements are not merely reactive to information accumulation but are learned over training to converge toward J-shaped trajectories and lead to more informative viewpoints. This directly increases task-relevant information and correlates with higher success rates. Notably, behavioral adaptations to sudden occlusions immediately upon their first presentation suggests that they can leverage a learned understanding of the task’s underlying structure to strategically sample information when faced with novel challenges.

Previous freely moving virtual reality platforms, such as FreemoVR (47) and related open-field systems (e.g. ⁴⁸), demonstrated that unrestrained rodents can interact with perspective-correct environments and exhibit naturalistic behaviors. More recent immersive head-mounted approaches, including MouseGoggles (49) and Moculus (50), have further expanded the available visual field through stereo or panoramic displays though they typically constrain translational locomotion. The freely-moving virtual reality system developed here bridges these approaches and enables quantitative investigation of a range of visual processing in which fully manipulable input is dependent on an animal’s location or movement. This includes aspects of 3D perspective, motion parallax, and figure-ground segregation. By accurately coupling the visual scene to the animal’s pose and movements, the system reproduces realistic spatial and temporal dynamics of a physical environment, while maintaining experimental control of the rendered scene. It also provides a significant benefit over standard head-fixed virtual reality systems by allowing naturalistic head and eye movements – key aspects of active sensing, driving vestibular signals essential for the brain to incorporate information about self-movement (51).

A key next step is to resolve the neural dynamics that support this strategy, spanning the full sensorimotor loop from retina to motor output. Neural representations are shaped by the interaction between sensory input, motor commands, task structure (e.g., movement costs, reward contingencies), and intrinsic computational limits of neural circuits (52). Early visual areas are expected to reflect stimulus-driven signals, whereas downstream motor regions encode action plans. How these constraints are progressively integrated across the pathway to generate commitment and action, however, remains largely unknown – a systems-level question that our paradigm is uniquely positioned to address (53).

Normative frameworks such as Bounded Rationality propose that agents operate under representational constraints, limiting what information is retained and how it is encoded (54, 55). Different constraints should produce distinct signatures in both behavior and neural activity, including systematic patterns of choices and errors (56). By combining large-scale neural recordings with our task, it will be possible to link behavioral strategies to population coding principles, test competing models of representational limits, and determine how distributed circuits construct and transform latent variables that guide active sensing and decision-making.

Resource Availability

Lead contact.

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Mackenzie Weygandt Mathis (mackenzie.mathis@epfl.ch).

Materials availability.

This study did not generate new materials.

Data and code availability.

• The mouse behavioral data analyzed in this paper has been deposited at Zenodo (https://doi.org/10.5281/zenodo.19091270) and is publicly available as of the date of publication.

• All original code for this project has been deposited on GitHub (https://github.com/MMathisLab/FreelyMovingVR4Mice).

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Methods

Experimental Model and Study Participant details.

Behavioral data was collected from 10 C57BL/6J mice (wild-type) across three laboratories, respectively 6 in the MLAI (M-Lab of Adaptive Intelligence), 2 in the Tolias laboratory, and 2 at the Niell laboratory. All mice underwent the same behavioral protocol, including training, and the two test phases (dual and multi-occusion tasks). They were housed on a reversed 12-hour light-dark cycle from 9 am to 9 pm. All experiments took part within the dark phase of that cycle. Both male and female mice were used and were between the ages of 8 to 20 weeks (MLAI: 5 females, 1 male; Tolias laboratory: 2 males; Niell laboratory: 2 males). Mice were placed on a controlled water access protocol for training, where they were maintained above 90% of their original body weight and were given no less than 1.5 mL of water per day. These experiments were carried out using the same protocol in three laboratories: the Tolias laboratory, Baylor College of Medicine (Texas, USA) and Stanford University (CA, USA) under the BCM OLAW Assurance number D16–00475 and Stanford Animal Welfare Assurance number D16–00134, the MLAI, Campus Biotech - EPFL (Geneva, Switzerland) according to the Geneva canton license number GE367, and the Niell laboratory, University of Oregon (Oregon, USA) under the protocol number AUP-21–21.

Method Details.

Behavioral experiments.

The same experimental protocol was followed in all three laboratories that carried out the experiments (Figure S2A).

Habituation and behavioral training

Prior to habituation and training, WT mice were placed on a controlled water schedule in their home cage. They were then habituated to the experimenter for a period of 3 to 5 consecutive days. On the first day, the experimenter placed their hand in the cage for 15 to 20 minutes to allow the mouse to acclimate to the experimenter’s scent. On the second day, the experimenter gently picked up the mouse using a plastic tube and then returned it to the cage. This was repeated multiple times during 15–20 minute sessions. On the third day, the mice were moved from the plastic tube onto the experimenter’s hands multiple times, allowing them to freely explore the hands. If a mouse showed signs of stress at any stage, the day was repeated until the mouse was comfortable. Following this habituation period, mice were habituated to the virtual reality arena by allowing them to freely explore the box for 30 minutes. To encourage mice to associate lick ports with rewards, a small water reward of 5 µL was given whenever mice were in close proximity to the right or left lick ports. This water reward was manually triggered by the experimenter using the manual water task on the vr4mice GUI.

The virtual world is rendered in the real-world arena on a 520-mm-wide screen. Once habituated, mice were trained in three phases: 1.1) detection without a velocity threshold, 1.2) detection with a velocity threshold, and 2) discrimination. Mice progressed through stages upon meeting a predefined criterion, which required completion of more than 125 trials in 60 minutes with a success rate greater than 70%. During the detection phase (1.), a single object, a white teardrop, appeared randomly on the right or left side of the virtual world with a probability of 0.5. The mouse reported the location of the object by moving to the lick port on the side of the object. In the first phase of the detection task (1.1), we did not enforce a strict trial initiation threshold: mice had to stay in the trial initiation area for at least 100 ms with a velocity lower than 20 Unity units per second (∼ 67.5 cm/s) to start the trial. Then in the second phase of the detection (1.2), the velocity threshold was reduced to 10 Unity units per second (∼ 33.75 cm/s), and the time limit increased to 250 ms. In the discrimination phase (2.), a distractor object, a black teardrop, was presented on the opposite side of the white teardrop, with still a probability of 0.5 to appear on one side or the other and a trial initiation threshold similar to 1.2.

Discrimination with occlusion

In the discrimination with occlusion test, occluding walls were placed in front of the objects of interest. These occluders consisted of two walls separated by a central aperture, allowing the mice to observe the objects through the gap. The extent of occlusion varied with the aperture width, causing the objects of interest to be either partially, fully visible, or fully occluded, varying the amount of information the mice could gather about the objects from a stationary position.

In the dual-occlusion test, mice observed objects through a wide or narrow aperture (Figure 3A), restricting the visible portion of objects of interest. When standing at the center of the trial initiation area, 99% of the objects was visible for the wide aperture condition (5.6 cm walls on each side, with an aperture of 40 cm), compared to 1% for the narrow one (19 cm walls on each side, with an aperture of 15 cm). The aperture condition was randomly selected for each trial using a uniform distribution.

After completing this task for five sessions, the mice were then tested on the multi-occlusion task, with five different aperture conditions ranging from 0% to 99% objects visibility on a logarithmic scale (0%, 10%, 35%, 72%, and 99% of the objects visible; Figure 5A), also from the center of the trial initiation area. This second task was also conducted for five sessions.

Inclusion criteria

To ensure that only sessions with sufficient task engagement were analyzed, we applied a post hoc inclusion criterion to test-day sessions (dual or multi-occlusion). Mice had to have completed more than 125 trials with a success rate greater than 70%, and their reward rate could not drop by more than 25% between the least and most occluded conditions for the session to be included. In total, across ten mice, we ended up with 43 sessions for the dual-occlusion test and 25 sessions on the multi-occlusion test (with one mouse that performed no session above the criterion). For the analysis correlating decision-point distance with success rate, all sessions with more than 125 trials and a total success rate greater than 70% were included regardless of the reward-drop criterion. This approach was taken to capture the full range of behavioral strategies across both high- and low-performing sessions.

The virtual reality system.

The arena

The virtual reality arena consisted of a transparent box measuring 52 × 52 cm. The arena was constructed using a base plate and four side panels, all made of PMMA (polymethyl methacrylate; Angst and Pfister). The base plate measured 520 × 520 × 5 mm, while each side panel measured 525 × 360 × 5 mm. These side panels were securely attached to the base using aquarium primer and glue (Coltogum). Inside the box, an additional floor plate with an anti-glare adhesive coating (Gardinia) was placed to minimize reflections within the arena. Sandwiched between the inner floor plate and the outer floor plate was a sheet of light diffuser (LEE filters), which allowed for even illumination of the arena. This inner floor plate could be easily removed for cleaning between sessions.

The mounting cage

The plastic box was mounted onto a cubic metal frame constructed from 25 mm aluminum optical rails (Thorlabs). The frame was built using eight 540 mm rails for the top and bottom faces of the cube and four 550 mm rails for the vertical sides. These rails were connected at the corners using Thorlabs corner cubes, with additional right-angled brackets added for extra stability. The transparent box was attached to this frame using custom 3D-printed top and bottom right-angled brackets. These brackets were secured to the optical rails with drop-in T-nuts and 1/4”−20 low-profile screws. The 3D-printed brackets also supported three computer monitors (SB241Y, Acer, 540 mm × 316.5 mm) coated with 0.9 neutral density filters (LEE filters) to achieve a maximum luminance of 10 cd/m².

Camera and computer

Above the cage, a high-speed USB camera (DMK - 37BUX28, Imaging Source) was mounted using a right-angled bracket attached to a cross-beam assembly consisting of one 540 mm beam and two 200 mm beams. The camera, equipped with a 3.5 mm EFL, F/1.4, 1/2” lens (Navitar), was positioned centrally over the arena. The cross-beam could be adjusted vertically, allowing the camera to have a complete field of view of the arena floor. The camera was connected to a Lambda vector Windows 10 workstation equipped with an Intel Core i9–10900X processor, 2 × RTX 3080 GPUs, 128 GB RAM, and two hard drives (1 TB for the operating system and 2 TB for data storage). This system managed camera capture, pose estimation, and generated the signals for water delivery.

Experimental environment acoustic stability

To ensure behavioral performance was not influenced by dynamical acoustic cues (e.g., electronic noise such as “coil whine” from the computer monitors due to changing brightness for instance), ambient sound was recorded during a representative 10-minute session using an iPhone 12 (Micro-Electro-Mechanical Systems microphone array; 44.1 kHz sampling rate) and analyzed post-hoc for spectral and temporal stationarity using the Librosa and SciPy libraries. The acoustic environment was characterized by a stable, low-frequency distribution. No periodic pulses, vertical spikes, or “blocks” of noise correlating with task-related screen updates was visible. The sharp horizontal cutoff at 16kHz represented the frequency response limit of the recording device (Figure S1H – Left-Top). The Root Mean Square (RMS) of sound pressure levels was extracted to monitor the noise for sustained step-functions or repeated task-locked pulses (Figure S1H – Left-Down). It revealed a highly consistent baseline noise floor. High-amplitude spikes correspond to mechanical artifacts, such as mouse or experimenter’s movement, rather than task-locked stimulus onsets. The absence of sustained amplitude shifts indicates that the computer hardware did not produce volume-based cues during task intervals. Power Spectral Density (PSD) was calculated for discrete 2.5-minute segments distributed across the session using Welch’s method to identify potential frequency shifts (Figure S1H – Right). The absence of unique narrow-band spectral peaks in any single segment confirms that no new dynamical frequencies (e.g., variable coil whine) were introduced during the execution of the behavioral task.

Water delivery and lick ports

Two lick ports were placed on the left and right sides of the arena, each located 5 cm behind the front screen. These lick ports consisted of custom-made 3D-printed boxes, each with a hole near the top to allow a length of water tubing to pass through. Inside each lick port, the water tube was securely fastened to the top of a steel pedestal (Thorlabs) using a separate 3D-printed attachment. This design ensured that the mice could not pull the tubing through the lick port and added enough weight to prevent the lick port from being knocked over by the mouse.

The water delivery system was controlled by a Teensy 4.0 microcontroller via a TTL (Transistor-Transistor-Logic) signal that was sent to the valve, with the duration of the TLL specifying the time in ms for the valve to stay open. Upon receiving a signal from the computer, the microcontroller would open one of the two solenoid valves, corresponding to either the right or left lick port. When a valve was opened, water would flow from a reservoir syringe (20 mL) through the tubing to the lick port, providing the mouse with a water reward of 3–5 µL.

To ensure that the valve’s open time matched the desired droplet volume, a calibration curve was created for the lick ports. The valve’s open time was adjusted in 100 ms increments, ranging from 100 ms to 1000 ms. At each interval, the droplet size was measured either by collecting the droplet in a weigh boat and weighing it on a precision scale or by using a pipette. These measurements were used to plot a calibration curve, which identified the time range that produced a droplet volume of 3–5 µL.

Prior to each experimental session the water systems tubing was cleared of any air bubbles by inserting a syringe plunger into the syringe reservoirs. Air bubbles were then pushed through the system by giving 5 × 1000 ms duration pulses, using the manual water task in the vr4mice GUI. Following this the plungers were removed and the syringes were topped up to 20 mL, ensuring the same water pressure between sessions.

Software.

DeepLabCut-Live! pose estimation and online kinematics

To provide closed-loop visual feedback to the mouse, its head position within the arena was estimated using the DeepLabCut-Live GUI (41), and used to render the corresponding viewpoint in Unity. DeepLabCut-Live utilizes a camera processor that receives images of the arena from the top camera at 100 Hz; complete parameters are listed under Camera and DeepLabCut-Live parameters used. The pose of the animal is inferred using the pre-trained TensorFlow SuperAnimal-TopViewMouse model (61) (Figure S1A), video adapted on a top-view video of our arena. This model, which has been optimized across multiple freely moving mouse datasets, yields 27 keypoints over the body of the mouse when viewed from above. From these keypoints, a custom-built DeepLabCut-Live! processor Python script computed kinematic features including the x- and y- position of the head, the running direction and the head angle relative to the body (head-body angle) (Figure S1B-C). The x- and y- position of the head were calculated by taking a weighted average of the eight keypoints corresponding to the head, namely head_midpoint, right_ear_tip, left_ear_tip, right_ear_base, left_ear_base, right_eye, left_eye, and nose:

$$W=\frac{{\sum }_{i=0}^7{w}_i{x}_i}{{\sum }_{i=0}^k{w}_i},$$ (1)

where $w_i$ is the confidence score of keypoint $i$ (i.e., DeepLabCut likelihood) and $x_i$ is the x-axis position of keypoint $i$ (same for y-axis). If the mean confidence of all head keypoints was below 0.6, the previous high-confidence position was used instead. This prevents low-confidence estimates from affecting head position in frames where the head is occluded, such as when grooming, or proximity to arena walls.

To compute the mouse running direction, we defined a body axis vector from the tail base keypoint to the neck keypoint. The running direction was the angle (in degree) between this body axis and the horizontal axis (x-axis). We then adjusted this angle so that 0° aligned with the main monitor, with values ranging from −180° to 180°. To obtain the head-body angle, we first calculated the head axis as the vector from the neck keypoint to the nose keypoint. The angle between the head and body axes then gave the head-body angle (Figure S1B). These kinematic features were sent to the vr4mice Python framework via a UDP (User Datagram Protocol) socket at a rate of around 70 Hz to determine when to trigger the trials.

Camera and DeepLabCut-Live parameters used.

Camera processor
camera_name	Imaging_source
resolution	(720, 540)
exposure	0.01
rotate	0
crop	(20, 530, 90, 620)
fps	100
display	True
display_resize	0.7
DeepLabCut-Live inference
model_name	superanimal_topview
model_type	base
cropping	None
dynamic_cropping	False, 0.5, 10
resize	1.0
mode	optimise_latency

VR4Mice Python framework

To run the virtual reality system, we created a custom Python framework called vr4mice. This framework acts as an interface between the incoming DeepLabCut-Live data, and Unity and controlled the valves for water delivery via a Teensy 4.0. In addition, the vr4mice software allows the user to edit task parameters (e.g., trial structure or task type) and save the data through a GUI.

vr4mice continually reads data from the DeepLabCut-Live GUI via the UDP socket on a separate thread and adds the incoming data to a buffer. The main thread periodically checks the buffer for the latest entry and reads it whenever Unity is ready to process a new frame. The data is then passed through a “one-euro” filter, a fast first-order low-pass filter with a cutoff that adapts to the sampling frequency, to reduce jitter in the signal (62). It is then converted from pixel space (720 × 540 pixels) to game space (18 × 8 game units) using the numpy interp function (63) and sent to the Unity game. The position in the arena, inferred from the DeepLabCut keypoints, is used to render the corresponding viewpoint and the running direction and head-body angle are used to initiate the trials in the game. At each iteration, the state of the mouse in the virtual game, which consists in its position and the reward status, is returned. If a reward was given in the Unity game, then vr4mice sends a signal via a serial port to a Teensy 4.0 to control the corresponding lick port.

The Unity game

The Unity game environment was constructed on a plane measuring 18 × 8 game units, which served as the floor of the virtual arena. The position of the real mouse’s head was transformed into the Unity coordinates and used to control a Unity empty object, representing the virtual mouse within the game, through the Unity Machine Learning Agents Toolkit, the ML-Agents package (64). A camera object was attached to this empty object, and its frustum was aligned with a projection plane (see the Off-axis projection Methods below for detail). This setup allowed objects of interest, designed with the software tool Blender (65) and randomly spawned behind the projection plane, to be viewed from the mouse’s perspective as it navigated the arena displayed on the main monitor.

We define a trial initiation area in the arena as a 24 × 18 cm (8 × 4 Unity units) zone centered on the x-axis. It is positioned 7.75 cm from the back wall and 21.25 cm from the screen. A trial initiates only when the mouse remains below a velocity threshold of 10 Unity units/s for at least 250 ms (20 Unity units/s for at least 100 ms at the start of training). As the game arena plane is not anisotropic, this corresponds to a velocity threshold of 33.75 cm/s on the y-axis and 15 cm/s on the x-axis, respectively 67.5 cm/s and 30 cm/s, considering that the mice direct themselves mostly on the y-axis at trial initiation. Additionally, the mouse’s head angle needed to be oriented toward the main monitor within a ± 90° angle to the perpendicular axis. This threshold, which does not account for the precise angle of the screen from the mouse position, was sufficient given the central placement of the start box. If those criteria were met, two objects of interest were spawned on the left or right of the projection plane on the screen. The mouse could then report the location of an object of interest by going to one of the two lick ports. To achieve this, two report boxes were implemented in the Unity game, with their positions corresponding to the lick ports’ location in the physical arena. If the virtual mouse stayed within the correct report box for at least 100 ms, a reward signal was sent to the vr4mice software via a socket. If the mouse reported incorrectly, no reward signal was sent. After the mouse had reported, the Objects of Interest disappeared from the projection plane and the next trial could be triggered by re-entering the trial initiation box.

Off-axis projection

In typical Virtual Reality systems, the subject’s visual field is centered within the projection screen and the virtual environment is rendered using an on-axis projection, where the camera frustum is symmetric on all sides relative to the center. The camera can be moved and rotated to follow the subject’s line of sight, maintaining a consistent and immersive experience. In our proposed system, the subject moves freely, changing the viewing angle relative to the screen. Consequently, an on-axis camera would fail to replicate certain aspects of natural vision, such as accurate motion parallax, leading to a less immersive experience. To address this, Cave Automatic Virtual Environment Virtual Reality (CAVE-VR) systems typically employ an off-axis projection camera instead (66). This approach adjusts the projection matrix to match the subject’s specific viewpoint, ensuring that virtual content is accurately rendered in alignment with the user’s perspective. It also maintains correct motion parallax, where distant objects appear to move less in response to head movements than closer objects, enhancing the realism of the rendering and making objects appear to be 3-dimensional.

To implement the off-axis projection, a projection plane was added to the Unity game. This projection plane acted as the near-clipping plane for the camera, meaning that any objects between the mouse and the projection plane would not be rendered. The central point of the camera perspective was fixed in the center of this projection plane. The corners of the frustum were fixed to the corners of the projection plane. The camera could then be moved around in Unity game coordinates and its view point rendered from the mouse’s perspective (Figure 1B; Figure S1D).

Latency testing

To measure the latency of the virtual reality system, the latency of the full system was measured along with the latencies for the DeepLabCut-Live and Unity steps (Figure S1E:G). A binary signal was generated in the DeepLabCut-Live processor so that it generates a binary signal $\mathcal{S}$:

$$\mathcal{S}=\frac{\mathrm{sign}(\sin (f\cdot \pi \cdot t))+1}{2},$$ (2)

where $f$ is the pulse frequency and $t$, the current time. Setting $f$ to 5 Hz, the pulse of the binary signal $\mathcal{S}$ flips every 200 ms. This signal $\mathcal{S}$ was saved alongside the camera timestamp and the timestamp at which DeepLabCut-Live had completed frame processing. $\mathcal{S}$ was initiated 10 seconds after the Unity game started, ensuring that it had time to load before the signal started. It was then sent to the Unity game via a UDP socket and rendered as a flashing black-and-white square in the top right corner of one of the side screens. A photodiode attached to this screen recorded the rendered signal $R$ via a Teensy 4.0. This signal $R$ was sent back to the computer via USB to be read on a separate thread, time-stamped within the DeepLabCut-Live processor script at 1000 Hz and saved alongside the data for the originally generated signal $\mathcal{S}$.

To measure end-to-end latency (Figure S1F), the photodiode signal $R$ was first low-pass filtered at 60 Hz to remove high-frequency noise. This filtered signal was then normalized between 0 and 1 and binarized using a 0.5 threshold. Rising edge timestamps from this binarized signal $B$ were compared to those of the generated signal S and their difference corresponds to the system’s end-to-end latency (Figure S1G). To determine the latency introduced by DeepLabCut-Live – namely frame capture, pose estimation, and keypoint averaging – we measured the time difference between the timestamp recorded at frame capture and the timestamp after computing kinematic features. The latency introduced by Unity was then calculated by subtracting the DeepLabCut-Live latency from the end-to-end latency (Figure S1E).

Behavioral analysis.

Quantifying information content across the arena

The information available to the mouse depends on its position in the arena, and how much of the two target objects were visible from that position. Here we defined information available as the objective visual evidence available in the environment (the surface of both objects), without making assumptions about the mice’s specific internal heuristic to solve the task (i.e., detection of the target object, or inference from the distractor object). For simplicity, we assumed these targets were circular. Each discrete point in the arena is treated as a visual sensor situated at a fixed elevation. The position of the target and aperture define the set of rays from the sensor that reach either target without intersecting the occluding walls. To calculate the unoccluded area of the targets visible from a given position, we compute the azimuthal angles subtended by unobstructed rays from each lattice point to both targets. These angular spans are converted to the corresponding circular-segment areas via $A(\theta )=\frac{1}{2}{r}^2(\theta -\sin \theta )$. The scalar information available at each position $(x,y)$ was then related to the visible areas as $I_{raw}(x,y)=\left(\frac{A_L}{d_L}+\frac{A_R}{d_R}\right)$, where $A_L$ and $A_R$ denote the visible segment areas of the left and right targets respectively and $d_L$ and $d_R$ are their corresponding Euclidean distances to the sensor. To faciliate comparison across conditions, the resulting information maps were min-max normalized to the range [0,1] across all conditions: $I(x,y)=\frac{I_{raw}(x,y)-I_{min}}{I_{max}-I_{min}}$,, where $I_{min}$ and $I_{max}$ are the minimum and maximum values of $I_{raw}$ across all information maps.

Off-line computation of kinematic and trial features

To analyze the DeepLabCut-Live data offline, keypoint traces were preprocessed to obtain traces that were free of keypoint jumps. This was achieved by removing all keypoints with a confidence value of less than 0.6. These low-likelihood keypoints were then inferred using neighboring keypoints using a linear interpolation. To reduce high-frequency jitter in keypoint tracking, the interpolated traces were smoothed using a Savitzky-Golay filter (SciPy (67)). To synchronize these traces with the Unity game data, the game data were first resampled to 50 Hz to correct for timing inconsistencies between frames. The DeepLabCut keypoint traces were then aligned by matching each timepoint to the closest one in the resampled Unity game recording. Considering that the Unity arena of dimensions 18 × 8 Unity units corresponds to a physical arena of 52 × 52 cm, and that screens are positioned 1 cm away from each wall, we interpolated the position of the animal with respect to the screens from the Unity coordinates to a physical space of 54 × 54 cm. We derived kinematic features such as the mouse’s head position, running direction and head angle relative to the body axis (head-body angle) similarly to what was implemented online for closed-loop feedback (Figure S1B). Running direction provides us with information on the spatial orientation of the mouse across the trial progression. Head-body angle serves as a measure of how attentional focus aligns with body movement, as supported by the activity of head-direction cells that are encoding the animal’s heading independent of body orientation (68). We also report the cosine and sine values for each angle to be used in the regression models, avoiding discontinuities and preserving directional structure. Cosine values are related to the forward/backward direction while sine values are related to the left/right direction in the arena. The x- and y-velocity in cm/s are computed from deriving the x- and y-position and used to obtain the combined velocity in cm/s. The initial position distribution heatmaps represent the spatial probability density of the animals’ starting coordinates, binned into a 3×3 grid. Each bin represents the mean occupancy for sessions and mice. We use this as an indicator of how consistent the mice were when approaching the start zone and how similar the initial stimuli were from trial to trial. We also tested for potential choice-predictive spatial biases by comparing the frequency of starts between Left and Right choices.

We also compute some trial metrics. To investigate whether previous trial outcomes influenced current choice behavior, we analyzed the side bias and decision stickiness. The side bias index detects persistent preferences for the left or right reward ports, while the decision stickiness index measures a potential tendency to repeat or alternate choices regardless of the outcome. Trial tortuosity is the trial path length (sum of the distances between all consecutive points) divided by the direct path from start to end position of the trial (distance between start and end points). From this, we defined change-of-mind trials as the 90^th percentile most tortuous forward trials. We classified a trial as “forward” if the minimum y-position was the initial position, hence excluding trajectories where the animal went toward the back of the arena during the trial. J-shaped trial rate is the rate of characteristic J-shaped trials, compared to wandering trials. We only include J-shaped trials, selected based on a duration (≤5 sec) and tortuosity (≤5) threshold, in our analysis, as an indication of the mouse’s engagement in the task. Time to report is the duration of the trial without the inter-trial interval, from trial initiation to the mouse getting to one reward port.

Training phase analysis

As mice required a variable number of days to complete training (Figure 2C), we assessed behavioral changes by selecting three representative sessions for each training phase (detection and discrimination). In the detection phase, the first and last of the detection without the velocity threshold and the last session of the detection with the velocity threshold. For the discrimination stage, the first, the penultimate, and the final session. These timepoints were chosen to approximate equivalent learning stages across mice. In the discrimination phase, some mice completed the phase in only two sessions, in which case no middle session was available. To account for this, statistical analyses were performed using a one-way ANOVA followed by Tukey’s HSD post-hoc test, excluding the middle timepoint when not present.

Trajectory analysis

For each session, we first preprocessed the data by 1) removing the inter-trial interval, and 2) removing the wandering trials to only keep the J-shaped trials. To characterize trajectories across occlusion conditions, the x- and y-position from individual J-shaped trials were first interpolated such that each trajectory comprised an equal number of time points (100 timepoints; trial progression). For each session, mean trajectories were computed by averaging the x and y coordinates across trials at each interpolated time point, separately for each aperture condition and choice direction (left or right). These session-level trajectories were then aggregated across mice to generate mean trajectories for each aperture. A similar procedure was used to calculate the other kinematics showed as a function of the trial progression.

For statistical comparisons, trajectories, running direction and head-body angles from right-choice trials were transformed by reflecting their x coordinates or angle across the midline (i.e., $x\to -x$), aligning them with the left-choice reference frame.

Trajectories approximation with curves of constant $L_p$ norm.

To approximate the trial trajectories, each interpolated trial was fitted with a curve of constant ${\text{L}}_p$ norm, which is a special case of the superellipse (Lamé curve) with equal horizontal and vertical scaling:

$${\left(\frac{|x{|}^p}{a^p}+\frac{|y{|}^p}{b^p}\right)}^{\frac{1}{p}}=1,a=b>0$$ (3)

where $p\ne 0$ is the shape exponent that controls the curvature of the curve and $a$ and $b$ are scaling factors set to the maximum Euclidean distance per trial. Higher values of $p$ produce more angular paths, axes-aligned paths that extend forward in the arena before turning toward the water ports, while lower values yield smoother, more direct trajectories, where mice move more directly to the water ports. For each trajectory, the optimal $p$ parameter was estimated by fitting the ${\text{L}}_p$ curve to the trials’ x and y position. So that they are comparable across trials, trajectories were first normalized by their maximum Euclidean norm:

$$\left({\tilde{x}}_i,{\tilde{y}}_i\right)=\frac{\left(x_i,y_i\right)}{{\max }_j\sqrt{x_i^2+y_i^2}},$$ (4)

where $i$ indexes points in the trajectory and $j$ spans all timepoints. The ${\text{L}}_p$ curve was the fitted by minimizing the discrepancy between the observed and predicted y-coordinates. Given a value of $p$, predicted y-values were computed from the normalized x-coordinates as:

$$\widehat{y_i}={\left(1-{\left|{\tilde{x}}_i\right|}^p\right)}^{\frac{1}{p}}.$$ (5)

The optimal $p$ was obtained by minimizing the sum of squared residuals:

$$L(p)=\sum _{i=1}^N{\left({\widehat{y}}_i-{\tilde{y}}_i\right)}^2,$$ (6)

where $N$ is the number of trajectory points. The resulting optimal $p$ values were averaged across trials to quantify trajectory shape as a function of occlusion condition.

Black target contrast experiment

To ensure that the strategy employed to perform the task is goal-directed rather than reflexive attraction to the high-contrast white teardrop or low-contrast black teardrop compared to the gray background, we performed the inverse contrast scheme experiment, i.e., the target object is the black teardrop and the distractor is the white one. All other parameters were kept the same. We trained three female mice, in the MLAI. It took an average of 10.3 ± 0.3 days and a total of 1489 ± 106.81 trials to train them. We tested each mouse on 6 days of the dual-occlusion task, and 6 days on the multi-occlusion task. 6, 5, and 4 sessions respectively were found above the inclusion criteria, for a total of 15 sessions for the dual-occlusion task (Figure S3Q). 4, 3 and 3 sessions respectively were found above the inclusion criteria, for a total of 10 sessions for the multi-occlusion task (Figure S5S).

Quantification and Statistical Analysis.

Statistical Analysis.

For trajectory and velocity analysis, a repeated measures two-way analysis of variance (ANOVA) was performed with aperture and y-bin location or trial progression as within-subject factors and session as the subject-level random effect (69). For the multi-occlusion task, variables for each aperture were compared using a repeated measures one-way ANOVA. Post hoc paired t-tests were conducted to further evaluate pairwise differences between aperture conditions. All resulting p-values were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) procedure (70).

Dynamic Choice Variable.

Logistic regression

We implemented a time-resolved multi-model framework to infer mouse choice from behavioral features using logistic regression (scikit-learn; (71)). Rather than fitting a single global model, we partitioned each trial into ten non-overlapping temporal segments, each representing 10% of the interpolated trial, and trained an independent model for each segment. This approach accounts for the non-stationarity of behavioral features by allowing regression weights to adapt to the specific dynamics of each trial phase.

For each of the ten models, we used a leave-one-session-out (LOSO) cross-validation approach, training one model per session on the data from N-1 sessions and evaluated on the single held-out session. Model performance was reported as the mean per-timepoint prediction accuracy across trials and sessions for the given time segment. To quantify the decision process, we also reported the model’s probability estimates for left versus right choices as a function of trial progression, which we considered as a dynamic decision variable. These estimates correspond to the predicted probability of a left choice (P(left) = 1 indicating full confidence in a left choice; P(left) = 0 for a right choice) at each timepoint. These probabilities were computed on the held-out validation session of each model and subsequently averaged across trials and sessions.

Bayesian Information Criterion analysis

To identify the feature set that best captures the animals’ underlying decision dynamics, we evaluated the Bayesian Information Criterion (BIC) (72, 73) for ten regressor sets, corresponding to distinct behavioral hypotheses. First, the Priors set, composed of the initial positions (x and y) and the trial history, tested whether the choice was pre-determined by starting conditions and past outcomes rather than current trial dynamics. Then a Kinematics set, comprising position (x and y) velocity (x and y), assessed the extent to which the decision was explainable from basic movement parameters. To further isolate the specific drivers of this movement, we tested two reduced hypotheses: Position only and Velocity only. These sets evaluated whether the decision variable was primarily reflected in the animal’s spatial coordinates or its instantaneous movement vectors, respectively. The Orientation set included the running direction and head-body axis (sine and cosine) to evaluate if the decision was represented in a purely egocentric or directional frame rather than a spatial one. The Lateral Displacement set solely comprised the features related to lateral movements (toward reward ports), namely x-position, x-velocity, and sine of head angle and running direction. The Task-relevant set expanded upon these by integrating kinematics and orientation with dynamic task features (trajectory tortuosity and length, and whether the trial was rewarded) to determine if the decision variable represented a synthesis of behavioral state and task context. Finally, the Full State set combined previously mentioned features, namely position, velocity, running direction, head-body angle, the dynamic task features and the prior features (initial position and trial history) to check whether pre-trial priors and online dynamics provide non-redundant contributions to the decision variable or if a more parsimonious model of current dynamics is sufficient to explain the behavior. The BIC is defined as

$$BIC=k\ln (n)-2\ln (\widehat{L}),$$ (7)

with $k$, the number of features used in the model, $n$, the number of observations (sample size) and $\widehat{L}$, the maximized likelihood. For our binary choice prediction, the log-likelihood is calculated as $\ln (\widehat{L})={\sum }_{i=1}^n\left[y_i\ln \left({\widehat{p}}_i\right)+\left(1-y_i\right)\ln \left(1-{\widehat{p}}_i\right)\right]$. This metric rewards models that assign high probability to the animal’s actual choices while penalizing excessive parameterization; consequently, lower BIC values indicate a more parsimonious fit.

However, rather than seeking a model based only on aggregated goodness of fit across timepoints and trials, we identified the feature set that most effectively reflects the dynamic evolution of the decision process as the trial progress. Hence, we evaluated three distinct criteria: mean accuracy, mean BIC score, and the Coefficient of Variation (CV) of the BIC within the final 10% of the trial $\text{(}C{V}_{end}\text{)}$. The later serves as a metric for terminal stability, defined as:

$$C{V}_{BIC}=\frac{{\sigma }_{BIC}}{{\mu }_{BIC}},$$ (8)

where ${\sigma }_{BIC}$ and ${\mu }_{BIC}$ are the standard deviation and lean of the BIC values calculated across individual trials, respectively. To integrate these metrics into a single selection criterion, we perfromed a min-max normalization for each component to a scale of [0,1]. We define a model efficiency score $S$, as the unweighted average of these normalized components:

$$S=\frac{Accurac{y}_{norm}+\left(1-BI{C}_{norm}\right)+\left(1-C{V}_{norm}\right)}{3}$$ (9)

This score identifies the feature set that maximizes predictive power (high accuracy) and parsimony (low BIC) while ensuring the model reaches a stable, low-variance state as the animal approaches its final choice (low $C{V}_{end}$). We found that the Lateral Displacement set reached the highest score when following those parameters, suggesting that the transition from medial movement toward the screen to lateral movement toward the ports most effectively capture the mice’s choice.

Decision point uncertainty threshold selection

To estimate the timing of decision commitment, we defined a decision point as the final timepoint in each trial where the model’s predicted probability crossed a specific threshold. To capture the turning point, we identified quantitatively the threshold that best captures the behavioral transition from screen-approach to port-directed movements. This ensures that the decision point used in downstream analysis aligns with the animal’s physical change.

We evaluated five candidate uncertainty thresholds $T\in \{0.1,0.2,0.3,0.4,0.5\}$. For each threshold, we calculated the change $(\text{Δ})$ in four kinematic variables by comparing the mean values of the 10 timepoints immediately preceding and following the predicted decision point: x-velocity (expected increase), y-velocity (expected decrease), absolute head-to-body angle (expected increase), and running direction relative to the screen (expected increase). To account for inherent trial noise, each kinematic variable $k$ was first Z-scored relative to a null distribution throughout the session, generated by randomly sampling two baseline windows within each trial:

$$Z_{k,T}=\frac{{\text{Δ}}_{k,T}-{\mu }_{k,null}}{{\sigma }_{k,null}}$$ (10)

To ensure that variables contribute equally to the selection process, we applied min-max standardization across all candidate thresholds, so that each Z-scored variable is in [0,1] while preserving the relative magnitude differences between thresholds:

$${\widehat{Z}}_{k,T}=\frac{Z_{k,T}-min\left(Z_k\right)}{max\left(Z_k\right)-min\left(Z_k\right)}$$ (11)

Finally, we calculated an aggregated turning score (S_T) for each threshold:

$$S_T={\widehat{Z}}_{\text{x}-\text{vel},T}-{\widehat{Z}}_{\text{y}-\text{vel},T}+{\widehat{Z}}_{\text{head angle},T}+{\widehat{Z}}_{\text{running dir},T}$$ (12)

Decision points analysis

Once we selected the threshold that best captured the turning point in the trajectory (i.e. 0.2 uncertainty threshold, 80% confidence level), we compared the spatial coordinate of these decision points across occlusion conditions, and more precisely their distance to the screen.

To investigate a potential link between the task performance and the spatial location of the decision points, we compared the average success rate to the average decision point location per mouse. In order to capture the full range of performances for that specific analysis, we included all sessions that were above 70% success rate for more than 125 trials, but that did not pass the reward drop criterion. First, we fitted a linear least-squares regression across mice and reported the Pearson correlation coefficient and p-value (Wald test, t-distributed test statistic), using scipy.stats.linregress. We also identified the upper and lower quintile groups of sessions with decision points closest to or farthest from the screen under the narrow aperture condition and compared average success rates between these groups between mice (all sessions for each mouse, not only the sessions contained in the quintile groups). Because all mice performed above the criterion (success rate above 0.7), this analysis provides a stringent test for a potential influence of strategy on task execution.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Experimental models: Organisms/strains
Mouse: C57BL/6J: Wild-type (WT)	The Jackson Laboratory	RRID:IMSR_JAX:000664
Deposited data
Mouse Behavioral Data	This paper	https://doi.org/10.5281/zenodo.1909127
Software and algorithms
Python Programming Language v3.9—3.10	Python Software Foundation (57)	RRID: SCR_008394; https://www.python.org/
DataJoint	Yatsenko et al. (58)	RRID:SCR_014543; https://www.datajoint.com/
DeepLabCut v2.2+	Mathis et al. (59), Nath et al. (60)	https://www.mackenziemathislab.org/deeplabcut
DeepLabCut-Live! v1.0.4	Kane et al. (41)	https://github.com/DeepLabCut/DeepLabCut-live
SuperAnimal-TopViewMouse	Ye et al. (61)	https://huggingface.co/mwmathis/DeepLabCutModelZoo-SuperAnimal-TopViewMouse
FreelyMovingVR4Mice	This paper	https://github.com/MMathisLab/FreelyMovingVR4Mice

Highlights.

• Mice perform active visual sensing in a free-range VR object discrimination task.

• They perform infotaxis and select the correct object under variable visibility.

• Decisions decoded from head-body movements and speed reflect occlusion sensitivity.