OPEN Dogs accurately track a moving object on a screen and anticipate its destination

Dogs are widely used in comparative cognition, yet their visual and motion perception is less understood. Anatomical and physiological evidence suggests high motion sensitivity in dogs (e.g., rod-dominant retina, high flicker-fusion rates), but behavioral findings are mixed. The study investigates whether dogs track and anticipate the destination of a moving object displayed on a screen. Using a naturalistic video of two people throwing a Frisbee, the authors ask to what extent the Frisbee’s motion explains dogs’ eye movements and whether dogs will anticipate the catcher’s location as the repetitive sequence unfolds (Experiment 1). They further test dogs’ responses to surprising events—freezing the scene and rewinding motion—to probe attention and prediction under violated expectations (Experiment 2). The work aims to establish anticipatory looking and motion tracking as tools for canine cognitive research and to clarify dogs’ predictive gaze control in dynamic scenes.

Prior work shows dogs’ motion processing differs from humans: dogs require higher coherence thresholds to detect coherent motion compared to humans and some other mammals. Evidence for sensitivity to biological motion in dogs is mixed, with inconsistent preferences for upright human or dog point-light walkers depending on view and stimulus configuration. Dogs interact preferentially with objects that display dependent (chasing-like) motion and habituate differently to such patterns, suggesting perception of animacy from motion. Eye-tracking research with dogs has mostly used static images, with a few studies using dynamic facial expressions or gaze-following to cues, but not focusing on continuous motion tracking or anticipatory looking. In humans, anticipatory looking is well documented during dynamic events (e.g., predicting ball bounces, grasp targets) and in infants for goals and object permanence; great apes also show anticipatory looks based on long-term memory and goal attribution. This background motivates testing whether dogs predict dynamic events and destinations in video stimuli.

Design: Two eye-tracking experiments presenting the same base video of two players throwing a violet Frisbee back and forth against a white background (10 throws).

Subjects: Experiment 1 tested 14 pet dogs (10 Border Collies, 3 mixed breeds, 1 Australian Shepherd; mean age 8.2 years; 9 females). Experiment 2 tested 12 pet dogs (8 Border Collies, 3 mixed breeds, 1 Australian Shepherd; mean age 7.6 years; 8 females). Inclusion required at least 70% of fixations within the video area; final analyzed samples: N=11 (Exp. 1) and N=9 (Exp. 2).

Stimuli and manipulations: Videos at 60 fps. Exp. 1 duration 16 s; Exp. 2 duration 24 s. Half the dogs viewed a mirrored version. Exp. 1: at four throws (2, 5, 7, 10), the catcher (player toward whom the Frisbee was moving) was frozen to assess whether catcher movement contributes to anticipatory looks. Exp. 2: the entire video was frozen while the Frisbee hovered midair (four times, 1000 ms each), then rewound until the Frisbee returned to the thrower; two such freeze-rewind sequences occurred around the 4th–5th throws and two around the 9th throw.

Apparatus: EyeLink 1000 (SR Research) sampled right-eye movements at 1000 Hz. Dogs’ heads stabilized with adjustable chin rest. Stimuli on 27-inch LCD monitor (1024×768; 60 Hz) at 50 cm; video subtended ~48.2°×30.9°. Frisbee diameter ~52 px (~2.6°). Calibration: 3- or 5-point with animated targets; central fixation (white expanding circle) triggered video when fixated 100 ms.

Procedure: After calibration, dogs viewed the video once per experiment. Data were mirrored back as needed to a common coordinate frame.

Measures and analysis: Focused on horizontal gaze. No event-detection; used raw gaze coordinates. Frisbee position was annotated as a dynamic AOI every ~61 ms (Exp. 1) and ~55 ms (Exp. 2). For each dog, linear models regressed horizontal gaze position on Frisbee x-coordinate to compute r² (variance explained). Anticipatory looking: for each catching event, determined the first time the dog looked at the catcher within ±650 ms of the frame before Frisbee-catcher contact; excluded the first throw and trials where the dog had not looked away from the catcher beforehand. Exp. 1 analyzed 10 catching events (throws 2–10; with catcher frozen in 2, 5, 7, 10). Exp. 2 analyzed 14 catching-like events (10 forward throws plus 4 rewound events) and five longer interest periods (first two and last two freeze-rewind sequences, and periods before, between, and after them). Linear mixed models (lme4) with subject as random effect tested predictors: Exp. 1 LMM 01: throw number (z-scored), catcher movement (moving vs frozen). Exp. 2 LMM 02: throw number (z), Frisbee movement (forward vs backward/rewind). Exp. 2 LMM 03: movement (forward vs backward) and interest period number within movement condition (z). Model diagnostics indicated no problematic collinearity or assumption violations; stability checks were performed.

Experiment 1 (N=11 analyzed):

• Motion tracking: Frisbee x-position explained a large portion of dogs’ horizontal gaze variance (median r²=0.61; range 0.01–0.89).
• Anticipatory looking: Dogs’ first look to the catcher occurred earlier with increasing throw number (LMM 01: χ²=4.81, df=1, p=0.028). Median latency across events centered around −54 ms (range −259 to +378 ms) relative to catcher-contact frame; by later throws, dogs typically looked before the Frisbee arrived. Catcher movement (moving vs frozen) did not significantly affect latency (χ²=0.17, df=1, p=0.678).

Experiment 2 (N=9 analyzed):

• Motion tracking: Frisbee explained gaze variance but less than in Exp. 1 (median r²=0.42; range 0.07–0.45).
• Attention to surprising events: The first two freeze–rewind sequences strongly captured attention (median r²≈0.76) compared to the last two sequences (median r²≈0.11) and to normal periods (beginning r²≈0.43; middle r²≈0.45; end r²≈0.49).
• Anticipatory looking with forward vs backward motion: Overall, neither throw number nor movement direction significantly affected latency in LMM 02 (throw: χ²=0.41, p=0.52; movement: χ²=1.15, p=0.28). Considering interest periods (LMM 03), dogs looked significantly faster to the catcher during forward motion than during backward (rewind) motion (χ²=4.22, df=1, p=0.040). Estimated mean latencies: forward ≈ −103 ms (95% CI [−232; −55]) vs backward ≈ 37 ms (95% CI [−59; 144]). Interest period number within condition was not significant (χ²=1.30, p=0.254). Occasional overshoot to the next catcher occurred at the first freeze in some dogs, indicating prediction beyond simple tracking.

Dogs accurately tracked a horizontally moving object on a screen and, with repeated exposure, shifted from reactive tracking to predictive gaze, anticipating the catcher’s location before contact. Anticipation emerged without reliance on the catcher’s own movements, indicating prediction based primarily on the object’s motion regularities. When the expected motion was disrupted (freeze and rewind), dogs initially devoted high attention to the Frisbee’s location, but this effect waned with repetition, suggesting transient surprise or weak gravity-related expectations and possible attenuation due to screen-based presentation. Dogs were faster to anticipate during normal forward motion than during backward (rewind) motion, demonstrating sensitivity to natural temporal directionality. The results support using anticipatory looking and motion tracking paradigms to probe predictive processing in canids and suggest that dogs’ gaze control can be guided by learned motion regularities. The discussion highlights practical considerations (e.g., frame/refresh rates near or above dogs’ flicker-fusion thresholds, stimulus parameters) and potential moderators such as breed differences and retinal specializations that could influence motion tracking along the horizontal axis.

The study shows that pet dogs can precisely follow a moving object on a screen and, with brief experience, anticipate its destination, evidencing predictive gaze control based on motion regularities. Anticipatory looking and motion tracking are promising, scalable methods for canine cognition research. Future work should: (1) optimize display parameters (e.g., >80 Hz refresh/frame rates, stimulus size, speed, luminance, contrast); (2) test different object identities and contexts (e.g., prey-like vs abstract motion); (3) examine breed-related visual specializations (e.g., visual streak differences); (4) probe physical expectations by freezing at different physical states (hovering vs post-impact) and manipulating temporal directionality (reversed playback); and (5) explore goal-based action prediction using interruptions within vs after action sequences.

• Display parameters: The 60 Hz frame/refresh rate may be suboptimal given dogs’ higher flicker-fusion thresholds; higher rates (>80 Hz) may improve perceived continuity and tracking.
• Sample composition and size: Small final samples (Exp. 1: N=11; Exp. 2: N=9) with many herding breeds may limit generalizability across breeds and ages.
• Screen-based stimuli: Reactions to surprising events may be attenuated on screens compared to real-world physics; ecological validity could be limited.
• Missing data: Reduced tracking during later freeze–rewind sequences led to missing observations for some latency analyses, potentially reducing power.
• Specificity of context: Only horizontal motion in a Frisbee-throwing context was tested; results may not generalize to different motion axes or object types.