STRIDE: Continual Learning using Skill Transfer and Reuse with Incremental Discrete Encoding

Developing general-purpose physics-based character controllers via reinforcement learning is fundamentally hindered by policy interference in continuous spaces. While training isolated controllers avoids interference, it incurs significant training overhead and yields disjointed behavioral transitions. To address these challenges, we present STRIDE (Skill Transfer and Reuse with Incremental Discrete Encoding), a framework for the continual learning and composition of diverse character behaviors. Instead of training independent models, STRIDE leverages policy quantization via a Vector Quantized Variational Autoencoder (VQ-VAE) to dynamically discover and encode reusable behavioral primitives within a discrete latent space. New behaviors and tasks are learned by combining existing codewords, providing a stable motion prior that structurally eliminates catastrophic forgetting. Evaluated across diverse locomotion styles, agile maneuvers like kicking, and composite tasks like target-aiming, STRIDE accelerates downstream learning by × 1.3 to × 2.0 compared to baseline learning frameworks. Furthermore, the discrete representations inherently improve policy interpretability, allowing direct tracking of skill activations and behavioral decisions during task execution. STRIDE offers a scalable, transparent solution for continual character control without compromising motion quality.