Facebook Trains AI-Powered Robot Dog with 70% Success

Researchers tested an AI-powered robot walking on different terrain. The robot created by a Chinese startup called Unitree looks like a four-legged animal that’s learning to walk across other surfaces. As the robot moves forward, it adjusts its step and adapts without falling. 

AI-powered Robot: Training

Facebook’s AI research team collaborates with the University of California and Carnegie Mellon University’s School of Computer to train the robot. As the three join forces, researchers taught the AI-powered robot to adjust to different conditions in real-time. 

In a video, the robot adjusts itself as it moves in different scenarios. It can walk over rocks, sand, a construction site, and even oil. The AI-powered robot can adapt as it encounters an elevation change. Yes, it can go up or down a step. 

Researchers placed a plastic across the room in a living room and poured oil, creating a slick surface. They even piled up planks and other obstacles to test the unnamed robot. It can even withstand when the researchers dropped the weight on the AI-powered robot’s back. Each time it’s tested, the robot maintains its balance and continues moving forward. 

A UC Berkeley Professor, Jitenda Malik, says the robot adapted quickly as they conducted the trial and error experiment. As the robot experiences various situations, it gathers information from its surroundings. The AI robot doesn’t have any computer vision. Through this process, it learns from how its body responds. 

Its process is almost similar to how humans learn. For example, if a person moves from hard to soft terrain, they will re-adjust their foot whenever it sinks. According to Malik, the “challenges of robotics” have many similarities with “real-world variability.”   

Computer Simulation Training

Researchers trained the AI-powered robot using a two-technique combination by exposing it to various surfaces and grueling conditions before testing it IRL. The team calls this AI-related breakthrough Rapid Motor Adaptation or RMA. It’s also the first to have an entirely learned-based system that enables a robot to adapt to its environment. The process is starting from scratch, then “by exploring and interacting with the world.”

This AI breakthrough could improve the performance of robots overall. AI-powered can use it in actual situations like search and rescue operations. Also, it can help at home where machines have to navigate changes in elevation like a ramp or stairs. The invaluable research can also be applied to “smart cities,” which use real-time data to help with traffic. It can also be helpful in other conditions that hinder a person’s quality of life. 

It’s possible to pre-program robots to navigate some environments. However, it’s tough for a programmer to predict every obstacle the robot will encounter. On the other hand, teaching a robot in real-time could work on cheaper hardware. So it can be helpful to keep the costs low in the future. 

Coronavirus Aftermath

Due to coronavirus, the AI researchers had to change how the experiments are conducted because the lab was closed. Ashish Kumar, UC Berkeley graduate, tested the robot in the following environments: 

• Inside the comforts of his home
• At a nearby construction site 
• On hiking trails in the Bay Area

Kumar elaborated by saying he tried conducting the experiment with whatever he could find in some sense. The robot also broke multiple times during the testing. The researchers compared how the AI-powered robot outperforms alternative systems. According to a summary, the robot can walkthrough: 

• Mud 
• Sand
• Trains 
• Dirt pile without the danger of falling

As a result, the experiment and training were successful in 70% of the trials. The majority of the trial was a success where it had to walk down a hiking trail. It did not fall 80% of the tests where it needed to walk through a pile of cement and a pile of pebbles.