Science China Information Sciences General Computer Science
Evolving embodied intelligence via reinforcement learning: challenges, barriers, and pathways
Abstract
Abstract Embodied artificial intelligence—which involves agents perceiving, reasoning, and acting within physical environments—represents a new frontier towards artificial general intelligence. This study explores the evolving landscape of embodied intelligence, and presents a comprehensive survey on leveraging reinforcement learning (RL) to enhance embodied intelligence, systematically addressing the evolution from single-task performance in controlled settings to sophisticated cross-environment and cross-task capabilities. The study delves into several virtuoso capabilities, such as morphological intelligence (MI), human-like values and ethics, spontaneous reward generation, and reflective equilibrium. These capabilities are essential for developing adaptable, ethical, and intelligent embodied agents capable of navigating complex real-world scenarios. Furthermore, the study identifies and analyzes critical barriers in embodied RL, including sample efficiency, generalization and transferability, safety and robustness, real-time learning and adaptation, and interpretability and transparency. For each of these barriers, we offer an overview of current approaches and propose potential avenues for future research.