TL;DR
While we can track multiple moving objects in our direct field of vision, our mental "simulation engine" appears strictly limited to a single item. Through a series of nine experiments, researchers demonstrate that imagining the trajectories of just two objects creates a massive serial bottleneck, forcing the brain to simulate one path to completion before starting the next.
What Problem Does This Paper Address?
Cognitive science has long known that humans have a "Multiple Object Tracking" (MOT) capacity in direct perception, allowing us to follow roughly 3 to 5 moving items at once. However, much of our reasoning about the physical world happens when objects are hidden or yet to move—a process called mental simulation.
The authors set out to determine the capacity limits of this "mind’s eye." If we can track four balls bouncing on a screen, can we also mentally simulate four balls bouncing in our heads? Previous research on static mental imagery suggested that adding objects increases difficulty, but the specific capacity for dynamic simulation remained a mystery.
Methods
The researchers developed a novel Imagined Objects Tracking task. Participants watched short 2D animations of balls moving (e.g., a cannonball arc or a horizontal slide). The animation would pause mid-motion, and participants were asked to mentally "continue" the trajectory and press a key the moment they believed the ball hit the ground or a wall.
The core of the study compared human performance against two computational models:
- Parallel Model: Simulates all objects simultaneously (predicting nearly identical response times for two balls hitting at the same time).
- Serial Model: Simulates one object fully, then "goes back" to simulate the second (predicting a significant time delay for the second response).
The researchers varied the "true impact time" across trials to see if participants' subjective timing scaled linearly with the physics engine's ground truth.
Datasets
The study relied on original data from nine pre-registered experiments involving a total of 313 participants.
- Experiments 1-5: Explored baseline tracking, perception controls, grouping cues, naturalistic occlusion, and simplified physics.
- Experiments S1-S3: Supplemental studies ruling out noise, lack of motivation (using monetary bonuses), and lack of practice as confounding factors.
- All raw data, modeling code, and stimuli are hosted on the Open Science Framework (OSF).
Key Insights and Ideas
- The Single-Object Bottleneck: In Experiment 1b (two objects), the data was a near-perfect match for the Serial Model. Participants didn't just lag; they appeared to finish simulating the first ball's entire path before even beginning the second.
- Imagination $\neq$ Perception: In a control experiment where the balls remained visible (Experiment 2), the serial delay vanished. This proves the bottleneck isn't a motor/response issue (the "fingers can't move fast enough" argument) but a fundamental constraint of the mental simulation process itself.
- Grouping is a "Hack," Not a Cure: Adding strong Gestalt cues (Experiment 3)—making both balls move in the same direction at the same speed—reduced the serial delay but did not eliminate it. This suggests we can use "heuristics" to group objects, but the underlying simulation engine still prefers a single-item stream.
- The Illusion of Vividness: The authors note a striking conflict: people feel like they can imagine complex, multi-object dynamic scenes vividly, yet their performance on timing tasks reveals they are actually "poor jugglers," processing the scene one piece at a time.
Key Results
- Model Fit: The Serial Model explained 96% of the variance in average response times for two-object tasks, while the Parallel Model explained only 20%.
- The "Serial Gap": When imagining two objects, the second response was delayed by an average of 640 ms compared to the first. In the perceptual control (where objects were visible), this delay dropped to just 88 ms.
- Linear Scaling: Even with the serial delay, participants' responses scaled linearly with the "true impact time," showing that they were accurately simulating the physics, just not in parallel.
- Naturalistic Generalization: The serial bottleneck persisted even when objects disappeared behind "natural" occluders (Experiment 4) rather than simply freezing in mid-air.
Limitations and Future Work
- Non-Interacting Objects: The study focused on independent objects. The authors acknowledge that if objects were to collide, a "serial" approach might fail, potentially forcing the brain to use an "interleaved" model (switching back and forth rapidly).
- Eye Movements: The experiments did not control for saccades. It is possible that participants "track" imagined objects by moving their eyes to specific coordinates, which is inherently a serial process.
- Feature Integration: Future work could investigate if shared features (like color or shape) allow for higher-capacity simulation through more advanced grouping mechanisms.
- Complexity vs. Capacity: While Experiment 5 used "hyper-simplified" physics, the limit remained at one object. This suggests the bottleneck is about the number of entities, not the mathematical complexity of the motion.
Summary updated