Researchers Turn a Mini Cheetah Quadruped Into a Highly-Successful Soccer Goalkeeper
Designed to split the challenging task into reusable skills, this framework gives Mini Cheetah a whole new level of agility.
Mini Cheetah, the quadrupedal robot platform from the Massachusetts Institute of Technology, has a new skill to show off thanks to researchers from the University of California, Berkeley: the ability to block nearly 90 percent of shots taken on a soccer goal with some impressive agility.
"Soccer goalkeeping using quadrupeds is a challenging problem," the researchers explain in the abstract to a paper brought to our attention by IEEE Spectrum, "that combines highly dynamic locomotion with precise and fast non-prehensile object (ball) manipulation. The robot needs to react to and intercept a potentially flying ball using dynamic locomotion maneuvers in a very short amount of time, usually less than one second."
To attack this problem — or, from a soccer perspective, defend — the team developed a hierarchical model-free reinforcement learning framework which splits the task into two distinct sub-tasks: low-level locomotion control to give the robot the agility it will need to intercept the ball; and high-level planning to pick the required skill and motion to stop the ball from entering the goal.
To prove the concept the team gave a Mini Cheetah some real-world testing, putting the robot in the goal of a "mini penalty field" designed to mimic part of a soccer pitch in miniature. Using an Intel RealSense D435i and a motion capture system to track the robot's performance, the Mini Cheetah was given the job of stopping the size-3 soccer ball from crossing the goal line as it was kicked or thrown at random speeds.
The results are impressive — not to mention fun to watch, as the Mini Cheetah springs up to intercept the shot on goal. The robot proved capable of stopping 87.5 percent of shots on goal, which puts it ahead of professional human goalies — albeit ones in a far bigger goal and facing considerably faster footballs.
"We showcase that the multi-skill RL framework significantly outperformed a model-based planner, and was able to adequately leverage the speciality of each skill," the team concludes. "In this work, we focused solely on the goalkeeping task, but the proposed framework can be extended to other scenarios, such as multi-skill soccer ball kicking."
The paper detailing the team's work is available on Cornell University's arXiv.org preprint server under open-access terms.