Disney plays with reinforcement learning, and the new robot looks like Star Wars

Disney’s new robot is here!

Without further ado, let’s first take a look at what it looks like –

Big twinkling eyes, a shaking head, flexibly swinging antennae, and a pair of short legs .

The walking is bumpy, but the key is still very stable:

Although it is a machine, it cannot Not to mention that this look is very "spiritual":

At the just-concluded International Robot Conference IROS, Disney Research Institute research scientist Moritz Bächer led a team to demonstrate this The new robot is so adorable.

Although this little guy doesn’t have an official name yet, he’s already making quite an impression with his performance

and common feet The biggest difference of the robot is that it can not only maintain dynamic balance while walking, but also express some "emotions" through movements.

According to the researchers, this is mainly because the reinforcement learning system that the team spent a year developing last year was used in the training process; and also We hired a professional animator to design the movements, and applied the animated characters' movements to express emotions to the robot.

Based on this, the little guy is no longer a cold machine. It is conceivable that if a GPT brain is installed on it, the dog at home will be in danger~~~

Risibai, I am a college student. I.

Many netizens thought of the characters in Star Wars after seeing it:

Another wave of netizens thought that the little guy was so "naive", It is clearly the real-life version of WALL-E (WALL-E, the character in "Wall-E"):

More netizens expressed the feeling of "standing in the world" "On the cusp of the robot revolution":

Why does this little guy walk so smoothly? How exactly did the developers give it "emotion"? Let's look down.

"Walking" in the Swiss Forest

The secret of shaking your head and moving smoothly lies in the design of the head and legs.

Among them, the little guy’s head is a four-degree-of-freedom design, which can realize tilted observation in four directions: up, down, left, and right:

And its legs have been increased to five degrees of freedom and have hip joints, so it can maintain dynamic balance when walking.

Whether you push it sneakily or deliberately drag the carpet under its feet, it can easily handle the challenge

The challenge is more The complex terrain makes it easy to walk in the wild forests of Switzerland:

It is steady yet charming, and you can even jump up and down while walking:

According to the introduction of the researchers, it actually took less than a year to achieve the effect from the concept to the final polishing

This is still Thanks to 3D printing technology and the use of modular hardware and actuators, the speed of design and iteration is greatly improved.

Let’s talk about giving it “emotion”.

Disney developed mechanical dolls as early as the 1960s and 1970s that moved in an emotional way

But as robots became more advanced and mobile, the design It becomes challenging to develop emotional behaviors that are compatible with it.

To this end, Disney used its trump card and hired animators who are good at making characters convey emotions through movement to design movements for the robots.

But when it comes to animators designing actions, they have to face a problem -

Animation tools usually do not have built-in physical properties, and the animators' vision is sometimes difficult to achieve in reality. This requires the team to spend a lot of time experimenting.

To bridge this gap, the team used a reinforcement learning system.

Behind is the blessing of reinforcement learning

The research team developed a process based on reinforcement learning, which combines the animator's visual vision with the robot's movements through

simulation.

This process essentially provides the animator with

physical constraints to ensure that the highly expressive actions designed by the animator can be implemented in reality. Or as close as a robot can do.

Morgan Pope, a scientist at Disney Research Institute, said:

This is not just walking. Walking is only one of the inputs to the reinforcement learning system. Another important input is how to walk. .

Fortunately, in this case, one of the advantages of reinforcement learning is that the generated motion is very robust.

And according to the researchers, this process can be used to train a robot to do new actions in just a few hours on a PC, greatly shortening the time required to develop new robot characters.

At the same time, during repeated training, subtle adjustments can be made to the motor performance, mass distribution, and friction between the robot and the ground.

Bächer, who is in charge of the team, believes that "this has shortened the time required for Disney to develop new robot characters from years to months"

And the researchers also emphasized the importance of this process sex, not the robot itself

If we want to add more legs, arms, or make a whole new character, we can quickly teach it new moves. We can apply off-the-shelf actuators, 3D printed components and adaptable reinforcement learning frameworks, all of which can be used for robots that look and act very differently. This robot is just the beginning of the journey.

#

The above is the detailed content of Disney plays with reinforcement learning, and the new robot looks like Star Wars. For more information, please follow other related articles on the PHP Chinese website!