A team of scientists from Stanford University and the Shanghai Qi Zhi Institute is dedicated to developing dynamic locomotion methods for a four-legged robot. They are using artificial intelligence (AI) algorithms, reinforcement learning and 3D depth cameras to do so. One of the key application areas for this technology is rescue missions. In these scenarios, robots must be able to independently identify and skillfully overcome obstacles.
The newly developed algorithms give a robot dog the ability to autonomously navigate and skillfully navigate obstacle-ridden terrain. This is achieved through adaptive control of the required motion sequences. The researchers plan to present these algorithms at the upcoming Conference on Robot Learning. They are open source and thus freely accessible.
The robotic dog is capable of assessing physical challenges and learning a wide range of movement skills and putting them into practice. Zipeng Fu, a participating scientist, provides the following explanation: “We combine perception and control using images from a depth camera and machine learning.”
Traditional learning methods are often based on complex reward systems that are specifically designed to deal with certain obstacles. However, this makes it difficult to adapt these methods to new environments. The team of scientists at Stanford is therefore taking a different approach, combining reinforcement learning with visual perception. To do this, they used the sensors and cameras already installed on the robot dog and developed a suitable algorithm.
Chelsea Finn, a professor of computer science at Stanford and lead author of the study, explains their method as follows: “We focus primarily on how far the robot can move forward and how much force it uses to do so. Over time, the robot then learns more complex motor skills that allow it to advance.”
During the testing phase, the robots demonstrated impressive capabilities. They were able to detect and overcome obstacles that were more than 1.5 times their own body height. They were able to jump over trenches that were wider than 1.5 times their body length. They were also able to crawl under obstacles that measured about three-quarters of their own body height and squeeze through narrow crevices narrower than themselves.
