SMT-Learner: Movement Trajectory Learning to Decode Motor Control Strategies

ICLR 2026 Conference Submission22147 Authors

Keywords: Autoencoders, Contrastive Learning, Transfer Learning, Movement Trajectory, Motor Skill Learning, Neuromotor Controls

TL;DR: Movement Trajectory Learning to Decode Motor Controls

Abstract: Spatiotemporal movement trajectory (SMT) representation is essential to understanding the motor skill learning and adaptation strategies that inform neurorehabilitation practices. Movement performance metrics (i.e., speed, accuracy) are insufficient to characterize motor control strategies and learning patterns, particularly in individuals with disordered movement. Motor skill learning patterns require an interpretable sequential SMT representation that preserves spatial, temporal, and performance variables. We present a novel SMT-Learner with transformer autoencoders that optimize performance-aware contrastive and adaptive transfer losses, combining cross-task and cross-subject transfer paradigms. SMT-Learner encodes trajectories into a high-dimensional latent space and enables motor performance-aware learning. We introduce an Exploration-Exploitation (E-E) analytical framework that quantifies motor skill learning and control strategies to balance different movement patterns and micro-adaptation. We tested and validated the SMT-Learner with two visuomotor reaching datasets: (1) a prospectively obtained cohort of term and preterm children's motor learning and performance of unimanual and bimanual tasks, and (2) extensively overtrained non-human primates performing target-directed reaching movements. Our ablation and baseline comparison across geometric, statistical, and clustering metrics demonstrated that SMT-Learner outperformed with the lowest reconstruction error (0.086) and optimized clinical correlation with motor performance variables. Investigated E-E patterns significantly correlated with the early and late stages of motor learning and speed-accuracy trade-offs principles. The SMT-Learner framework provides an efficient computational approach to quantify motor learning strategies; potential advanced downstream applications in developmental assessment, neurorehabilitation monitoring, and movement optimization in robotics or brain-computer interfacing.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 22147