Reinforced Intelligence Through Active Interaction in Real World: A Survey on Embodied AI

초록

Embodied AI is a transformative field that enables intelligent systems to actively interact with and adapt to complex physical environments. This survey examines recent advancements, focusing on how embodied agents bridge the gap between cyber-physical systems and human-centered settings. Key developments include leveraging foundation models for high-level planning, affordance learning, and low-level control, alongside unifying data from internet-scale, simulation, and real-world sources. Reinforcement learning (RL), inverse RL, and imitation learning have been pivotal in advancing robotic control and skill acquisition. Additionally, the transition from transformers to deep state-space models (SSM) offers new possibilities for enhancing prediction and decision-making capabilities in embodied systems. We also discuss challenges and future directions, highlighting the importance of generalization, affordance learning, and the path toward artificial general intelligence (AGI). This survey provides a concise roadmap for researchers and practitioners shaping the future of embodied intelligence in real-world applications.

키워드