Mechanisms for Autonomous Sub-goal Discovery in Lifelong Robotic Learning

Abstract

[Abstract] Lifelong learning in robotics seeks the continuous acquisition and reuse of knowledge to autonomously master complex tasks. A key challenge is discovering sub-goals without relying on explicit intermediate rewards, enabling robots to decompose tasks and transfer skills across domains. We propose a dual mechanism for sub-goal discovery. The first is a top-down strategy that builds hierarchical sub-goal chains from general goals via intrinsic motivations. The second is a bottom-up approach that uncovers latent links between previously learned goals and perceptual classes. Implemented in the e-MDB cognitive architecture, our method was tested in both simulation and a real-world robotic manipulation task. Results show efficient sub-goal generation, transfer, and generalization.