Goose Factory's robot dog takes over the "job" of real dogs! He can play games and play games happily, and he can also walk people around 6

Let the robot dog learn the movement data of the real dog, it really feels like a dog!

It jumped the hurdle easily, and the "owner" behind it almost failed to keep up:

It was easy to drill a "dog hole" :

You can also have two dogs having fun together, one chasing and the other escaping...

This is the latest progress of Goose Factory Robot Dog.

Using the pre-trained model to feed the robot dog the movement data of real dogs and through reinforcement learning, the robot dog Max is not only more agile in behavior, but also can adapt based on the skills it has mastered. A more complex environment.

To sum up, it’s a bit more doggy inside and out.

Chasing Sahuan'er strategically

Among the new skills that the robot dog learned this time, the strongest one is playing games.

The robot dog can not only abide by the rules, but also come up with its own strategies to win the game. It may be smarter than a real dog.

Specifically, this is an obstacle chasing game, inspired by "World Chase Tag", with the following rules:

The researchers set different game difficulties , the simplest is an open field:

During the game, the robot dog obviously has a strategy.

For example, usually the pursuer will launch a fierce attack until the dodger is far away from the chess flag, forcing it into a blind corner, and the game ends.

If the pursuer finds that the dodger is very close to the flag and has no chance to catch up with it, it will give up the pursuit first and wait for the next one The flag appears:

It doesn’t matter if there are obstacles, the two dogs play equally well 6:

And they can perform like this, and Not the robot dog was trained with this game from the beginning.

It is actually based on some actions, knowledge and skills that you have learned to deal with this game scene.

How to implement it specifically? Look down.

Learn the real dog data

The research is divided into three stages.

• Learning animal movement postures
• Connecting movement postures with external perception
• Additional network acquisition and information related to complex tasks

The first stage is to use the motion capture system commonly used in games to collect posture data of real dogs, including walking, running, jumping, standing and other actions, and build an imitation learning task in the simulator. The information in these data is then abstracted and compressed into a deep neural network model, so that it can cover motion posture information while also having a certain interpretability.

Tencent RoboticsX Robotics Laboratory cooperates with Tencent Games to use game technology to improve the accuracy and efficiency of the simulation engine. At the same time, it has accumulated a variety of motion capture materials during the game production and development process.

#These technologies and data also play a certain auxiliary role in the training of agents based on physical simulation and the deployment of real-world robot strategies.

In the process of imitation learning, the neural network only accepts the proprioceptive information of the robot dog as input

, such as the status of the motor on the robot, etc.

In the next step, the model introduces sensory data from the surrounding environment, such as obstacles underfoot that are "seen" through other sensors.

In the second stage, through additional network parameters, the animal posture mastered in the first stage is connected with external perception.

In this way, the robot dog can respond to the external environment through the actions it has learned.

When the robot can adapt to a variety of complex environments, the knowledge that connects the animal posture with external perception will also be solidified and stored in the neural network structure.

Then the robot dog can go up the stairs freely.

Or running on discontinuous or uneven ground:

Then it comes to the final stage, where the robot dog solves practical problems based on the skills learned above

This is the game-making process mentioned above.

According to reports, all control strategies for the robot dog in the game are neural network strategies.

Learn in simulation and through zero-shot transfer(zero-adjustment transfer), let the neural network simulate human reasoning to identify new things that have never been seen before, and transfer This knowledge is deployed on real robots.

For example, if you have learned how to avoid obstacles in the pre-trained model, then if you set up obstacles in the game, the robot dog can easily deal with them.

This new research progress is brought by Tencent Robotics X Robot Laboratory.

The experiment was established in 2018. The robot projects currently launched include the first/second generation robot dog Max, the robot dog Jamoca, the wheeled robot Ollie, self-balancing autonomous motorcycles, etc.

One More Thing

It is worth mentioning that scholars at UC Berkeley also used a "real dog" training method on robot dogs.

Pieter Abbeel, Ng’s founding disciple, and others let the robot dog roll on the ground for an hour and learned to walk.

When Tencent released the second generation of its robot dog Max last year, in a small tidbit, the dog could "flip its feet" and "run around". It does have the smell of a furry child at home.

(Of course, if you want it to become a dog that listens to its owner, you can give it orders through commands.)

Let’s just say that the current development direction of robot dogs is not to do somersaults, but to “grab jobs” with real dogs?

The above is the detailed content of Goose Factory's robot dog takes over the "job" of real dogs! He can play games and play games happily, and he can also walk people around 6. For more information, please follow other related articles on the PHP Chinese website!