Predicting motor intention in hand reaching from gaze dynamics

Understanding how perceptual information guides intentional action remains a fundamental challenge in cognitive neuroscience. Eye movements provide a non-invasive behavioral measure of this process, as gaze reflects both attentional selection and prospective motor planning. This thesis investigates whether and how gaze dynamics encode human intention during goal-directed hand-reaching actions and whether intended targets can be predicted before movement onset, with the help of an immersive virtual reality (VR) experimentation.Three VR-based hand-reaching experiments were conducted, with different levels of planning time and task demands, including fully free-choice reaching and constrained-based selection of single and multiple objects. Gaze and hand movements were recorded with high spatiotemporal resolution, analyzed and decoded using an epoch-based framework distinguishing early planning, late motor planning, and movement execution. The VR approach allowed a precise assessment of the temporal evolution of intention-related gaze signals in a fully controlled environment. Across these and other preliminary experiments, results revealed that gaze con-sistently preceded hand movement and contained predictive information about the intended targets, both before and after movement initiation. Decoding performance improved progressively within trials, with the lowest accuracy during early planning and the highest during late planning and execution. In delayed response conditions, gaze showed prolonged exploratory behavior before the go signal, followed by increased fixation stability on the selected target. Sequential reaching tasks additionally revealed hierarchical planning through anticipatory fixations on upcoming targets. Intention decoding was implemented using two approaches: an analytical rule-based baseline and a feedforward neural network operating on gaze velocity and fixation stability. Both methods predicted intended targets above chance level, with only marginal performance gains from the neural network.