Researchers from Facebook and US universities have unveiled a robotic dog that works on a machine learning model and is ready for ‘mishaps’ such as impassable surfaces, variable heights, and heavy weight.
Unless you’ve been living under a rock for the past few years, you’ve likely seen at least one video of the Boston Dynamics Robo-Dog Spot now owned by Hyundai Corporation. If you are also a devoted geek reader You saw one of these in action right here in Israel, with Israeli technology.. Now Facebook robotics and artificial intelligence researchers are taking and updating the concept with their very own robotic dog that knows how to deal with real-world conditions and is armed with some particularly advanced models.
The AI that will improve robots?
A team of researchers from Facebook’s AI division and the American universities of Berkeley and Carnegie Mellon have introduced a new artificial intelligence model called Rapid Motor Adaption (or RMA for short) designed to improve the mobility of robots. The model allows robots to make corrections to their movement in real time, under different conditions and under different circumstances.
The model created by the research teams for the robots is based on the use of two very familiar techniques from the world of AI and machine learning: the first is reinforcement learning (or RL) and the second is supervised learning. Using these techniques, the researchers created a situation in which the robot becomes accustomed to changes in real time, such as the surfaces on which it walks or carries weight without warning, without the help of visual feedback. In other words, while Spot the Dog scans its surroundings with computer vision, Facebook’s bot is ready for failure and quickly adapts to the UAV.
The researchers behind the new model point out that today robots are either manually programmed according to the environment in which they are designed to operate, or they are partially programmed manually, using learning techniques to learn to navigate the environment. On the contrary, RMA, according to the researchers, is the first model based solely on learning techniques that allows robots to adapt to different environments from scratch by moving in space and interacting with the environment.
According to the researchers, the robots they ran the RMA on were able to achieve more successful results than competing systems when it comes to walking on different surfaces, including different gradients and obstacles, and carrying different weights on them as they walk. “It is even more difficult than sophisticated manual programming, because it is difficult or impossible to pre-program a robot to get used to the full range of environments in the world,” the researchers wrote.
Prepare the robot for real life
Investigators who signed Article On the subject of the model developed together with Facebook, they state that no matter how good the teams that develop robots are, their full or partial programming will always be successful under laboratory conditions, but will not hold up in real life testing. According to them, only the use of a model like RMA could give robots the ability to move in space while loading different weights, without the need to repair their software each time; Or the ability to keep walking properly even if they have suffered some damage to one of their “feet” and the ability to adapt to countless changes that can occur in real time.
To address the various challenges of space movement and real-time repairs, RMA relies on two subsystems. The first is a basic policy created by RL-based learning simulations: the researchers stored a lot of information about the different environments (such as the amount of friction in each environment or the different weights thrown at the robot) and from this information they learned to anticipate which settings must do.
Treat your telegram to the largest technology channel in the country
Join the Gictic channel on Telegram
However, in real time it is impossible to know exactly which surfaces the robot will encounter, and this is where the second subsystem comes into play: the adaptation module. The basic policy is the one that drives the robot in real time and is designed to operate quickly, otherwise it will freeze in place or crash. Next to it, the adaptation module runs in the background, which takes the information collected from the robot’s sensors and makes the necessary corrections according to this information. The two subsystems operate asynchronously, which also allows for a smaller computational module needed to run RMA, so that in practice there is less weight on the robot.
Using all of these abilities, crews were able to march their RMA-based robot into environments such as sand, mud, walking paths, high meadows, and even a pile of dirt when only one of the experiments failed the mission. Among other things, the robot managed 70% of attempts to go down high stairs, which were found to walk where it walked, and 80% of attempts to walk on piles of gravel and concrete. The robot was also able to walk at 12 kilograms, the same weight as its entire body weight. Additionally, the Robo-Dog was able to steadily walk on an oiled surface which caused the dog without the model to instantly slip.
The complete study is Here.