Robot Parkour Learning

Reproducing teacher-student policy learning for quadruped robot parkour — AIR, Tsinghua University

Demonstration of the teacher policy for quadruped robot parkour.

Overview

During my internship at the Institute for AI Industry Research (AIR), Tsinghua University (Dec 2025 — Mar 2026), I worked on reproducing and improving Robot Parkour Learning, a work by Ziwen Zhuang, Zipeng Fu et al. published at CoRL 2023 (Oral, Best Systems Paper Award Finalist).

The project uses an end-to-end reinforcement learning framework that enables low-cost quadruped robots (e.g., Unitree Go1/Go2) to perform parkour skills — climbing over high obstacles, leaping across gaps, crawling under low barriers, and squeezing through narrow gap.

My Work

  • Reproduced the three-stage training pipeline: soft-constraint pretraining → hard-constraint fine-tuning → DAgger knowledge distillation
  • Trained the teacher policy with RL in simulation, achieving skills such as climbing 0.40m obstacles and jumping 0.60m gaps
  • Waiting for a Chance to deploy the distilled vision-based policy on Unitree Go2 with onboard Nvidia Jetson + Intel Realsense depth camera
  • Analyzed the teacher-student performance gap and proposed improvements for sim-to-real transfer

References