This study presents prompt-Adaptive Continual Learning (PACL), a novel framework that combines prompt-based learning and reinforcement learning to enable adaptive robotic task execution in dynamic, non-stationary environments. PACL addresses the challenge of catastrophic forgetting through a dual-memory architecture that separates historical and new experiences. It uses real-time sensory input to generate context-aware prompts, which guide a decision-making model optimized via reinforcement learning. Regularization techniques help preserve previously learned behaviors, maintaining performance across tasks. The architecture consists of a prompt generator and a decision-making model trained in tandem for continual adaptation. Empirical results from simulated robotics environments show improved task adjustment efficiency and robustness. By integrating prompt engineering, PACL enhances interpretability and scalability, making it suitable for real-world applications with shifting task demands. This approach offers a memory-efficient and flexible solution that advances continual learning in robotics by balancing adaptability and stability.
Fan et al. (Fri,) studied this question.