Sankar, R.; Suryawanshi, A.; Rougier, N. P.; Leblois, A.

The acquisition of sensorimotor skills critically depends on basal ganglia (BG)-thalamo-cortical circuits. Prevailing theories propose that the BG optimize motor output through reinforcement learning (RL), using internal performance evaluations to approximate stochastic gradient ascent. However, this framework struggles in non-convex performance landscapes, where local optima hinder efficient learning. Songbirds provide a compelling biological example of robust sensorimotor learning, mastering complex vocalizations through trial-and-error within a specialized BG-thalamo-cortical architecture. Here, we present a computational model constrained by the anatomy, physiology, and developmental trajectory of the zebra finch song system. The model combines a BG-driven RL pathway with a parallel cortical motor pathway that progressively consolidates successful motor patterns via Hebbian plasticity. In addition, we incorporate synaptic volatility within the BG pathway, introducing structured variability across learning. Through simulations of vocal learning using both a biophysical syrinx model and synthetic performance landscapes, we demonstrate that this dual-pathway architecture reliably converges to global optima and outperforms standard and noise-annealed RL approaches. The model reproduces key experimental features of song learning, including non-monotonic learning trajectories, a gradual reduction in motor variability, and the developmental transfer of motor control from subcortical to cortical circuits. Mechanistically, delayed maturation of the cortical pathway provides an implicit regulation of the exploration-exploitation trade-off, while synaptic volatility enables escape from local optima. These results highlight the importance of neural circuit architecture and dynamics in efficient learning, and suggest biologically inspired design principles for improving the robustness and sample efficiency of artificial RL systems in complex sensorimotor domains.