A robotic goalkeeper with four legs that uses reinforcement learning

A robotic goalkeeper with four legs:

A reinforcement learning algorithm has recently been developed by scientists from the Hybrid Robotics Group at UC Berkeley, Simon Fraser University, and the Georgia Institute of Technology that enables a quadrupedal robot to play soccer effectively in the position of goalkeeper. The model, which was described in a study that was pre-published on arXiv, gradually develops the robot’s skills through trial and error.

According to the study’s authors, Xiaoyu Huang, Zhongyu Li, Yanzhen Xiang, Yiming Ni, Yufeng Chi, Yunhao Li, Lizhi Yang, Xue Bin Peng, and Koushil Sreenath, “by allowing quadrupeds play soccer, we can test the limits of the artificial intelligence of athletic legged robots.” “Goalkeeping is a fascinating but difficult activity that calls for the robot to respond to the quick-moving ball, often flying through the air, and intercept it using dynamic moves in a very brief amount of time (usually within one second). By figuring this out, we can also learn more about how to build dynamic, intelligent legged robots.”

The main goal of Huang and his colleagues’ latest work was to develop a four-legged robot goalkeeper that can develop its abilities as it plays, exactly like a human goalkeeper would. To achieve this, the researchers used a reinforcement learning model that trains the robot through a process of trial and error rather than using a predetermined, human-engineered technique.

According to the researchers, “the robot first learns several locomotion control policies to perform diverse skills, such as sidestep, dive, and jump, while tracking randomized trajectories for the robot’s toes.” “After analyzing the detected ball position and the robot’s states, the robot learns a high-level planning policy based on these control rules to determine an ideal skill and motion to intercept the ball.”

Through several soccer simulations, the researchers educated their reinforcement learning algorithm. They then tested its performance in the real world by deploying the rules it had learnt on the Mini Cheetah, a genuine quadrupedal robot created at the Massachusetts Institute of Technology (MIT).

The Mini Cheetah robot’s abilities as a soccer goalkeeper were shown to be significantly enhanced by the reinforcement learning framework developed by Huang and his colleagues. In the group’s actual tests, the robot was successful in stopping 87.5% of 40 random gunfire.

The quadrupedal robot Mini Cheetah is able to conduct very dynamic and agile locomotion abilities, such as jumping and diving, as well as rapid and precise manipulation skills, such pushing the ball away using its swinging legs, utilizing our proposed strategy, the researchers added. This truly stretches the limits of movement on two legs by demonstrating that the leg can function as a manipulator, much like a human’s can.

Future robots intended to compete in the RoboCup and other robotic soccer events may perform better thanks to the reinforcement learning model developed by this study team. Additionally, their model might

The researchers continued, “We anticipate that we can soon make quadrupedal robots competitive with human soccer players.” In order to play soccer, robots “need to perform a broader variety of dynamic and agile motions and gain higher intelligence.”

Read More

• How to Build a 3-phase Distribution Panel
• Japan introduces delivery robots that are humble and friendly.
• Researchers make new solar cell material more stable.