A quadruped robot learns to "stand on two legs and walk down stairs"! 83% more efficient than legged systems

​Remember that robot racing with Tesla?

This is a four-legged wheel-legged robot with the same name as the company developed by a spin-off company from the Swiss Federal Institute of Technology in Zurich—— Swiss-Mile, formerly known as ANYmal quadruped robot.

Less than half a year since it raced with Tesla, it has achieved another major upgrade.

##This upgrade has improved the robot’s algorithm, and its movement ability has directly improved UP UP UP!

You can stand on both legs and go down the stairs:

(Editor’s inner OS: If I wear roller skates and go down the stairs, I may fall A dog eats shit)

If you are tired of climbing the stairs, take the elevator and open the elevator door with your front foot:

##Faced with obstacles:

##It also knows when to Stand up, when to "lay down", the switch between standing legs and four-legged movement is smoother:

AMP algorithm is applied to real robots for the first time

Swiss-Mile has previously used model predictive control (MPC) and reinforcement learning (RL) methods, however this requires tedious adjustments Only in this way can you get the ideal exercise method.

In this algorithm upgrade, the researchers used Multi-AMP (Adversarial Motion Priors) The algorithm enhances traditional reinforcement learning frameworks to automate the imitation target and motion selection process for multiple motion priors without heuristics.

What exactly is AMP?

This is an adversarial learning system based on physical character animation. It was proposed by researchers at the University of California, Berkeley and Shanghai Jiao Tong University, and Swiss-Mile applied this method to it for the first time. On a real robot!

For general imitation learning, it is usually necessary to manually extract a large number of motion clips that need to be imitated as tracking targets. However, using AMP can automatically select appropriate motion clips to achieve the target task. It outsources error measurement, phase and motion clip selection to Discriminator,The discriminator learns to distinguish between policies,and state transitions of motion data.

Simulation and deployment training

Researchers will use the Multi-AMP framework Deployed on Swiss-Mile with 16 degrees of freedom and implemented using the Isaac Gym simulator, more than 4000 robots can train skills simultaneously in 42 minutes.

The training environment consists of three tasks:

The first one The task is quadrupedal locomotion, and the motion data consists of movements recorded by the RL strategy.

The second task is to avoid the skill, allowing the robot to hide under the table. The skill's motion data is generated by the trajectory optimization pipeline and deployed and tracked by the MPC controller.

The last task is the movement conversion between "standing" and "four-legged". The robot uses data to solve The Coupling Skill allows you to stand up on your hind legs, slide on both legs, and finally sit down again using the same motion as when standing up.

Finally, Swiss-Mile was deployed into a real environment. The researchers used an actuator model of the leg joint to Bridge the gap from simulation to reality, and use rough terrain training, random interference, etc. to improve robustness; if a certain joint speed exceeds the limit of the actuator, the robot learns to maintain a safe tolerance of the limit through the terminal trajectory.

Quadruped or humanoid robot? 83% more efficient than legged systems!

Swiss-Mile is not only a four-legged robot, but also a humanoid robot.

The wheeled type has many significant advantages over the legged type. After algorithm improvements, the robot can stand up directly in the state of a "humanoid robot" and perform gliding and climbing stairs. , downhill and other difficult movements, it can move faster and more efficiently, much faster than walking on four legs, and 83% more efficient than the legged system!

In the future, the company hopes to commercialize wheel-legged robots to complete a variety of tasks, including mapping, inspection, Disaster relief and logistics in urban environments, etc.

Maybe one day, you will see the robot using its raised "front legs" as arms to grab express packages, then put them in the cargo compartment on its back, and then return Get down on all fours and get it delivered to your door as quickly as possible. ​

The above is the detailed content of A quadruped robot learns to "stand on two legs and walk down stairs"! 83% more efficient than legged systems. For more information, please follow other related articles on the PHP Chinese website!