Robots teach themselves to walk for an average of 3.5 hours with Google’s new algorithm

In the field of robotics, keeping robots standing and moving smoothly has always been a tricky challenge, as it requires a high level of expertise and design. Although some traditional robots can stand and move under artificial control, their range of activities is also fraught with limitations.

Robots teach themselves to walk for an average of 3.5 hours with Google's new algorithm

Pictured is Rainbow Dash moving on an empty doormat

To solve this problem, Google recently published a paper with researchers at the Georgia Institute of Technology and the University of California, Berkeley, detailing how they successfully built a robot that could walk on its own through AI. They gave the four-legged robot a lovely code name, “Rainbow Dash.”

According to the world record, the fastest speed for a baby to walk from crawling to learning to walk is six months, and according to the test data in the paper, Rainbow Dash takes an average of about 3.5 hours to learn to move forward, backward and left and right – 1.5 hours on hard, flat ground, the robot learns to walk. It takes about 5.5 hours on a mattress made of memory sponge and about 4.5 hours on a hollow carpet.

Specifically, the robot uses deep-enhanced learning, which combines deep learning with enhanced learning in two different types of AI techniques. Deep learning allows the system to process and evaluate raw input data from the environment in which it is located, and by intensive learning, algorithms can experiment to learn how to perform tasks and receive rewards and penalties based on the degree of completion. In other words, in this way, the robot can implement automatic control policies in an environment that it does not understand.

In previous experiments, researchers initially asked robots to simulate real-world environments. In the simulation environment, the robot’s virtual body first interacts with the virtual environment, and then the algorithm receives the virtual data until the system has the ability to “handle” the data, and a machine person with the physical form of the system is placed in the real environment for experimentation. This method helps to avoid damage to the robot and its surroundings during trial and error.

However, while the environment is easy to model, it is often time-consuming and the real world is full of unexpected situations, so there is limited significance in training robots in a simulated environment. After all, the ultimate goal of such research is to prepare robots for real-world scenarios.

Researchers at Google and the Georgia Institute of Technology and the University of California, Berkeley, are not “old-fashioned.” In their experiments, Rainbow Dash was trained from the start in a real-world environment, so that robots were able to adapt well not only to their environment, but also to similar environments.

Although Rainbow Dash is able to move independently, this does not mean that researchers can “let go” of it. At the very beginning of learning to walk in an environment, researchers still need to manually intervene hundreds of times on Rainbow Dash. To solve this problem, the researchers limited the environment in which the robot moves and allowed it to perform multiple motion training at once.

After the Rainbow Dash self-taught walking, the researchers were able to control the robot’s desired motion trajectory by connecting the control handle stoin, keeping the robot in a set environment. In addition, the robot automatically walks back when it recognizes the boundaries of the environment. Outside a particular environment, a robot may repeatedly fall and cause damage to the machine, and another hard-coded algorithm is needed to help the robot stand up.

Jan Tan, Google’s director of the study, told the media that the study took about a year to complete. He said:

We are interested in enabling robots to move in complex real-world environments. However, it is difficult to design a motion controller that can handle diversity and complexity flexibly.

Next, the researchers hope their algorithms will be suitable for different kinds of robots, or for multiple robots to learn in the same environment at the same time. Researchers believe that cracking the robot’s ability to move will be key to unlocking more practical robots – humans walk with their legs, and if they don’t use them, they can’t walk in the human world.

However, allowing robots to walk in the human world is a crucial task, and they can instead explore different terrains or unexploited areas of the earth, such as space. But because the robot relies on an action capture system mounted on it to determine its location, the device is not yet ready for use directly in the real world.