P‐4.28: Reinforcement Learning‐Based Embodied Intelligence Training
TL;DR
A systematic review of reinforcement learning ‐ based simulation training frameworks for embodied intelligence focuses on the fundamental principles of reinforcement learning, domain randomization strategies, and their roles in facilitating sim‐to‐real transfer.
Abstract
Embodied intelligence emphasizes the ability of an agent to learn and make decisions through continuous interaction with its environment, with a core principle being the tight coupling of perception, action, and cognition. In recent years, reinforcement learning (RL) has emerged as a key enabling paradigm for embodied intelligence due to its strong performance in sequential decision‐making tasks. However, training policies directly in real‐ world physical environments is often costly, risky, and inefficient. To address these challenges, high‐fidelity simulation environments are widely adopted for policy pretraining, while techniques such as domain randomization and sim‐to‐real transfer are employed to bridge the gap between simulation and reality, commonly referred to as the reality gap. This paper presents a systematic review of reinforcement learning ‐ based simulation training frameworks for embodied intelligence. We focus on the fundamental principles of reinforcement learning, domain randomization strategies, and their roles in facilitating sim‐to‐real transfer. Finally, we discuss current limitations and outline promising directions for future research.