Are you serious? Let the robot dog be the goalkeeper and publish a paper

Is it reliable to let a robot dog serve as a football goalkeeper? Whether it is reliable or not, let’s see the effect first and then draw a conclusion.

The staff made a very mild attack, and the robot dog blocked the ball:

More difficult, It’s no problem to score a parabolic goal:

Is it suspected of cheating to throw the ball by hand? Try it with your feet, and the robot dog can also defend the goal.

Interestingly, the study also tried to use a robot dog to Play football, and the other one will be the goalkeeper. It’s great that the two robot dogs can play by themselves:

After seeing the effect, I feel like the robot dog is the goalkeeper. Pretty reliable. This robot dog is Mini Cheetah, which was developed by MIT in 2019. Now researchers from the University of California, Berkeley and other institutions have deployed a new reinforcement learning framework for Mini Cheetah, allowing it to complete football goalkeeping tasks with a goalkeeping success rate of 87.5%. .

Paper address: https://arxiv.org/pdf/2210.04435.pdf

Kick the ball 4 meters away

Mini Cheetah successfully kept goal in less than 1 second

It is still difficult for Mini Cheetah to learn to keep goal, because it involves the height and dynamics of the object (such as the ball) being thrown Moving position, specifically, one side manipulates a fast-moving ball whose direction and position are uncertain, while the other side needs to quickly judge the ball's position to prevent a goal. Accomplishing this requires teaching the robot to dynamically move its body while ensuring that its feet (or face) get to where they need to be to block the ball in time, which is basically two puzzles rolled into one.

The solution of this research is to combine the motion controller with the end effector trajectory planning, so that the best way can be found to make Mini Cheetah before the ball reaches the target. Within one second, block.

To complete the above process, Mini Cheetah also needs to be trained to master a set of useful goalkeeping skills. For example, Mini Cheetah needs to master sideways interception of the ball near and near the ground, and master diving. Technique to reach the lower corner of the goal, jump to the top and upper corner of the goal. After completing these actions, Mini Cheetah can recover and finally land safely. Reference movements for each skill are manually programmed, trained in simulation, and then transferred directly to the robot.

The goal defended by Mini Cheetah is 1.5m wide and 0.9m high, the ball (number 3) is kicked from about 4m away, the ball is tracked externally and then Mini Cheetah blocks it. The performance of such a small robot dog to complete the ball-blocking action is impressive.

The research shows that the robot dog system can transfer the dynamic movements and goalkeeping skills learned in simulation to a real quadruped. On the robot, the goalkeeping success rate against random shots in the real world was 87.5%. The average success rate for human soccer goalkeepers is 69%. The researchers say their proposed framework can be extended to other scenarios, such as multi-skill football.

Let’s take a look at the framework behind this robot dog.

Hierarchical Reinforcement Learning Framework

First of all, making a four-legged robot a football goalkeeper is a very challenging problem, because it must simultaneously solve the problem of predicting the trajectory of an object and capturing non-grabbing objects. Two practical problems with holding objects (spheres). The robot needs to react to and intercept a ball flying in the air within a very short period of time (usually less than a second).

To accomplish this challenge, the research team proposed a hierarchical model-free reinforcement learning (RL) framework. The framework contains a multiple control strategy for different motor skills, covering different areas of the target.

These control strategies allow the robot to track the trajectories of randomly parameterized end effectors while performing specific motor skills such as jumping to block the ball, diving The ball and the ball are stopped rolling on the ground.

#The RL framework contains a high-level planner that helps the robot determine the required locomotor skills and plan the end-effector trajectory to intercept the fly direction Balls in different target areas.

This research deployed the above RL framework on the Mini Cheetah quadruped robot proposed by MIT in 2019. Experiments show that this RL framework can allow the quadruped robot to effectively intercept fast movements in the real world. ball.

Previous research on the RL framework of quadruped robots mainly focused on low-level motion control, such as making the robot walk at a required speed and imitating reference motion. The framework proposed in this study extends the learned motor skills to higher-level tasks, successfully using advanced planning to allow a quadruped robot to accurately intercept a fast-moving football with agile movements. This has important implications for advanced planning control of quadruped robots.

The above is the detailed content of Are you serious? Let the robot dog be the goalkeeper and publish a paper. For more information, please follow other related articles on the PHP Chinese website!