Reproduce “Learning Highly Dynamic Skills Transition for Quadruped Jumping Through Constrained Space” on OpenResearch

About this paper

Zeren Luo, Jiahui Zhang, Yimin Han, Ji Ma, Minghao Lu, Ioannis Havoutis, Peng Lu

Researchers from the University of Hong Kong and the University of Oxford developed ConsJump, a hierarchical reinforcement learning framework that enables a quadrupedal robot to autonomously perform highly dynamic jumps through narrow gates. The system integrates low-level imitation learning for agile gaits with a high-level perception-aware controller, demonstrating robust adaptation to varying gate positions and complex terrains on a real Unitree Aliengo robot.

Key insights

• Decoupling low-level skill acquisition (via imitation learning) from high-level task decision-making significantly reduces training complexity and enhances stability compared to end-to-end approaches.
• Leveraging a pre-trained low-level policy allows the high-level controller to explore a smaller, semantically meaningful action space (velocity commands), leading to higher task rewards and greater stability.
• The seamless integration of vision-based perception for real-time environmental awareness with hierarchical control enables dynamic, informed decision-making crucial for autonomous navigation in constrained environments.

Summary by alphaXiv.