Towards Quadrupedal Jumping and Walking for Dynamic Locomotion using Reinforcement Learning

2025-10-28

AI Summary

This work addresses the challenge of synergistic control for dynamic jumping and robust locomotion in quadrupedal robots. We propose a curriculum-based reinforcement learning framework that operates without predefined reference trajectories. Our method innovatively incorporates projectile motion physics to densify sparse rewards and introduces a reference-state initialization mechanism to enhance policy exploration efficiency. Integrating dynamics modeling, phase-wise reward shaping, and Sim2Real transfer techniques, we achieve efficient deployment of learned policies onto the real-world platform “Olympus.” Experiments demonstrate vertical jumps up to 1.0 m, horizontal jumps up to 1.25 m, omnidirectional jumping capability, centimeter-level positioning accuracy, and seamless transition between jumping and walking on unstructured terrain. Key contributions include: (i) a physics-informed reward construction method; (ii) a generalizable initialization strategy for omnidirectional dynamic locomotion; and (iii) a unified control paradigm enabling high-performance jumping and adaptive walking.

Technology Category

Intelligent Robots: Motion and Path PlanningSearch and Optimization: Learning to SearchHumans and AI: Human-Aware Planning and Behavior Prediction

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingResponsible Web: Machine-in-the-loop, human agency and autonomySemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

This paper presents a curriculum-based reinforcement learning framework for training precise and high-performance jumping policies for the robot `Olympus'. Separate policies are developed for vertical and horizontal jumps, leveraging a simple yet effective strategy. First, we densify the inherently sparse jumping reward using the laws of projectile motion. Next, a reference state initialization scheme is employed to accelerate the exploration of dynamic jumping behaviors without reliance on reference trajectories. We also present a walking policy that, when combined with the jumping policies, unlocks versatile and dynamic locomotion capabilities. Comprehensive testing validates walking on varied terrain surfaces and jumping performance that exceeds previous works, effectively crossing the Sim2Real gap. Experimental validation demonstrates horizontal jumps up to 1.25 m with centimeter accuracy and vertical jumps up to 1.0 m. Additionally, we show that with only minor modifications, the proposed method can be used to learn omnidirectional jumping.

Problem

• Training quadruped robots for precise jumping using reinforcement learning
• Developing combined walking and jumping for dynamic locomotion
• Bridging simulation-to-reality gap for versatile robotic movement

Innovation

• Curriculum-based reinforcement learning for quadrupedal robot jumping
• Densified sparse rewards using projectile motion laws
• Reference state initialization without trajectory dependency