CMU robot dog, standing upside down and going downstairs! Release is open source
Plus double height jump and fully autonomous parkour
To be honest, I’ve seen a lot of robot dogs doing crazy things——
But I was still surprised today.
The latest results from CMU directly teach dogs:
High jump twice as long as your body , long jump, stand on your head and even walk down the stairs upside down.
Not much to say, let’s just show the picture and experience it:
△ This is long jump
△ This is a high jump
△ Stand on your head and have fun
△ Go down the stairs upside down
I have to say, especially the "struggle" in the high jump section makes the dog particularly soulful.
Going up a ridge, crossing a gap, or crossing a slope is called slipping.
Even if there are "mistakes" in the middle, it will not affect its immediate progress.
After hearing this, Mr. LeCun had to give him a thumbs up.
How to refine such a soul?
In the tweet, the author analyzed this dog’s technology one by one.
After all, any mistake could be fatal.
In this regard, CMU uses sim2real to achieve precise foot control and challenge, maximizing mechanical advantages.
Among them, the simulator uses Gym.
Second, stand on your head. Walking on two legs is obviously much more difficult than walking on four.
But CMU's robot dog uses the same basic approach to achieve both tasks at the same time, and can even walk down stairs while standing on its head.
Third, for parkour operations (the focus of this study), the robot dog must decide its own direction through precise "eye muscle" coordination, rather than obeying human commands .
For example, when crossing two slopes in a row, it needs to jump up the slope at a very specific angle and then immediately change direction.
In order to learn these correct directions, CMU uses the MTS (Mixed Teacher Student) system to teach the robot dog.
Among them, only when the predicted direction is close to the true value, it will be adopted by the system.
Specifically, the system is divided into two stages :
In the first stage, RL is used to learn a movement strategy. This process can access some privileged information. In addition to environmental parameters and scan points (scandots), CMU also provides some appropriate landmark points (waypoints) for the robot dog for the purpose of guiding General direction.
Then, Regularized Online Adaptation (ROA) is used to train the evaluator to recover environmental information from the observation history.
In the second stage, the strategy is extracted from the scandots. The system will autonomously decide how to move forward based on the strategy and depth information, thus outputting motor commands in an agile manner.
The whole process is like "the teacher teaches, and the students learn by analogy."
In addition to this system, since parkour requires a variety of different actions to cross obstacles, it is also a headache to design a specific reward function for each obstacle.
Here, the author chose to formulate a unified and simple inner product reward function for all tasks.
Without it, the dog would behave like this:
Finally, CMU also proposed a new dual distillation method for extracting agile movement instructions and rapidly fluctuating forward directions from depth images.
Similarly, without it, the dog behaves like a "drunkard":
After the above operations, this dog finally learned a new way of autonomous parkour, interspersed with difficult moves.
Isn’t it exciting? Don't worry:
At the same time, the paper is also online. You can get it at the end.
about the author
Two of them co-authored the work, and both are Chinese :
One is Xuxin Cheng. This work was completed when he was a graduate student at CMU. He is now a doctoral student at the University of California, San Diego (UCSD), and his supervisor is Wang Xiaolong;
The other one is Shi Kexin, a visiting scholar at the CMU Robotics Institute. She graduated from Xi'an Jiaotong University with a bachelor's degree.
Project homepage (including links to papers, code, etc.): https://extreme-parkour.github.io/