Core Concepts
This study showcases a curriculum-based approach to deep reinforcement learning for quadrupedal jumping, eliminating the need for reference trajectories and achieving impressive results in real-world experiments.
Abstract
The study explores a novel curriculum-based reinforcement learning approach for quadrupedal jumping skills. It demonstrates the ability to learn dynamic jumping without relying on imitation or pre-existing reference trajectories. By leveraging a curriculum design, the study achieves versatile omnidirectional jumping motions, including forward jumps, diagonal jumps, and overcoming obstacles. The proposed method outperforms existing literature by achieving a 90cm forward jump on similar robots. Additionally, the robot exhibits continuous jumping capabilities on soft grassy terrains not included in the training stage. The study introduces domain randomization to bridge the simulation-to-real gap successfully. Experimental validations include various types of jumps, such as forward, diagonal, and continuous jumps, showcasing robust performance across different terrains and scenarios.
Stats
Particularly we achieve a 90cm forward jump, exceeding all previous records for similar robots reported in the existing literature.
Using 4096 agents and 24 environmental steps per agent per update step.
The three highly parallelized training stages took approximately 1.4 hours, 4.1 hours and 4.8 hours respectively.
We performed all of the experiments on the Unitree Go1 robot.
Real animals exhibit a four-legged contact phase during long-distance jumps.
Quotes
"Learning dynamic locomotion is still an open challenge despite recent works."
"Deep reinforcement learning has shown impressive generalization capabilities in executing locomotion tasks."
"Our approach eliminates reliance on pre-computed motion references."