This four-legged robot learned parkour to better navigate obstacles

Key Points:

  • ANYmal robot upgraded to perform parkour moves
  • ETH Zurich team enhances proprioception of ANYmal for better movement
  • Swiss team combines machine learning with model-based control for robot skills


ETH Zurich researchers have enhanced their four-legged robot, ANYmal, enabling it to perform basic parkour moves such as jumping across gaps and climbing obstacles. The robot, originally introduced in 2019 with reinforcement learning, has now improved proprioception to sense movement better.


In previous experiments, a team led by Nikita Rudin showcased a trio of customized ANYmal robots designed for planetary exploration, each with specific functions and backup capabilities. The robots were equipped with lidar sensors and specialized tools for tasks like surveying, chemical identification, and general measurements in challenging terrains resembling lunar and Martian landscapes.


Although ANYmal’s parkour abilities are a significant advancement, the robot is not yet as agile as humans due to inherent challenges in dynamic motion control and decision-making in real time. However, the Swiss team’s approach combines machine learning with model-based control, incorporating perception, locomotion, and navigation modules to help ANYmal adapt to varying terrains and obstacles.


Researchers used machine learning to teach ANYmal new skills like climbing and jumping, while model-based control enabled the robot to identify and navigate through obstacles more effectively. Tasks such as jumping and climbing were successfully executed, showcasing the robot’s ability to adapt its movements to diverse scenarios.


While ANYmal’s performance in controlled settings like box-jumping and obstacle negotiation is promising, the scalability and adaptability of its skills to more complex and unstructured environments remain to be fully tested. The team acknowledges that further development is needed to address the challenges of real-world navigation and improve response times in unpredictable scenarios.



Prompt Engineering Guides



©2024 The Horizon