Document Type

Unpublished Paper

Publication Date

2025

First Page

1

Last Page

16

Abstract

This survey reviews recent advances in applying reinforcement learning (RL) to enable dynamic and ballistic motions in legged robots, including running, jumping, stair climbing, and parkour. Focusing on high-agility behaviors that challenge traditional control frameworks, we categorize foundational locomotion tasks and highlight the RL methods, such as Proximal Policy Optimization, curriculum learning, and hybrid model-based strategies that have proven effective. We discuss the key challenges in transferring learned policies to real-world robots, managing uncertainty, and integrating perception and proprioception. Drawing from over 150 recent works, we provide a structured taxonomy of objectives, algorithms, and platforms, and identify trends in simulation frameworks and deployment strategies. This review aims to serve as a resource for researchers developing next-generation legged robots that can achieve robust, adaptive, and high-performance motion in complex environments. We provide a companion repository that maps the collection of papers, their methods, tasks, and platforms. This repository can be found at: https://github.com/DIRECTLab/legged_review

Download

Share

COinS