Abstract
ABSTRACTPredictable Failure over Uncertain Success: VIP Extrapolation Study
The Visual Inverted Pendulum (VIP) task, in which participants stabilize a visually animated unstable inverted pendulum with a joystick, is a reductionist model for studying postural control. This paradigm minimizes low-level sensory reflexes to isolate the role of higher-level processes during continuous balancing. The driven pendulum's deterministic yet unpredictable behavior necessitates constant intervention amid inherent sensorimotor delays, implying that predictive mechanisms are critical for proficiency. This dissertation investigates the development of predictive internal models – internal representations of the VIP's dynamics and the joystick's response properties – and the behavioral strategies these models enable. In Chapter One, I introduce the background literature that led to my hypothesis that such models are honed through error-based learning to forecast key kinematic landmarks along the pendulum's trajectory.In Chapter Two, I describe the use of inverted pendulum simulations to characterize its behavior and identify scenarios suited for probing predictive capabilities. To quantify the quality of stabilizing joystick maneuvers, further simulations generated scenario-specific solution manifolds by decomposing joystick maneuvers into magnitude, onset, and duration parameters. This parameterization allows profiling the parameter combinations participants select before and after training, thereby characterizing strategic shifts as they learn to predict pendulum outcomes.
To effectively utilize these scenarios, a novel approach transformed the continuous VIP task into a series of discrete, repeatable trials. As detailed in Chapter Three, these scenarios were designed around the pendulum’s separatrix with varying difficulty levels to enhance sensitivity to changes in predictive accuracy. Participants completed these probe trials before and after VIP training and had two tasks. In each trial, they viewed a briefly visible pendulum trajectory and were required to first predict its outcome as a passive observer and subsequently on a separate block of trials, generate a joystick maneuver to stabilize it at its direction of balance.
In Chapter Four, I tested the hypotheses that, following VIP training, participants would demonstrate: 1) The speed and accuracy of perceived IP destiny in probe trials will improve, 2) The speed and accuracy of motor adjustments of IP destiny will improve, and 3) the optimality of motor adjustments of IP destiny in probe trials will improve.
The results for perceptual judgments strongly aligned with predictions, showing significant – though differential – improvements in speed and accuracy across scenarios. In contrast, motor adjustments yielded a more complex pattern. While their initiation became considerably faster, accuracy improvements were confined to lower-difficulty scenarios. Furthermore, post-training motor optimality revealed that for high-difficulty scenarios, adjustments deviated further from the solution manifold, whereas for lower difficulties, parameter shifts suggested a strategic sensitivity to trade-offs between certainty and performance.
Collectively, these findings demonstrate that discrete VIP scenarios enable focused investigation of predictive control, and that predictive mechanisms honed through error-based learning are essential for sustained stabilization. The dissociation between refined perception and more complex motor strategic shifts reveals how users navigate time constraints and uncertainty. Implications for this work and future directions are discussed.