Although acrobatic robot displays seem like a wonderful marketing gimmick, they are usually well planned and expertly rehearsed. A four-legged AI robot has now been trained by researchers to navigate challenging, never-before-seen obstacle courses in practical settings.
The real world's inherent complexity, the quantity of information robots can gather about it, and the speed at which judgments must be made to perform dynamic motions make building nimble robots difficult.
Organizations such as Boston Dynamics have frequently published films of their robots performing a variety of tasks, including parkour and dancing. Even while these achievements are astounding, they usually require people to laboriously program each step or repeatedly practice in extremely controlled situations, reports Edd Gent in Singularity Hub.
The ability to apply abilities in the real world is severely limited by this approach. However, using machine learning, researchers from ETH Zurich in Switzerland have taught their robot dog, ANYmal, a set of fundamental locomotive skills. With these skills learned, the dog can now navigate a wide range of difficult obstacle courses both indoors and outdoors at up to 4.5 miles per hour.
After segmenting the problem into three sections, the researchers allocated a neural network to each part in order to develop a system that was both adaptable and capable. Initially, they developed a perception module that builds an image of the terrain and any obstructions in it using data from lidar and cameras.
Related Robot Completes Surgery in Space
They integrated this with a locomotion module that has picked up a wide range of abilities, such as jumping, climbing, crouching, and climbing down, to help it get over various barriers. Ultimately, these modules were combined with a navigation module that could determine which abilities to use to overcome each obstacle and plot a path through a sequence of them.
Instead of using human examples throughout the training process, the researchers exclusively used reinforcement learning, or trial and error. This allowed them to train the AI model on a huge number of randomized scenarios without having to manually label each one.
The fact that the robot uses chips that are implanted within it rather than relying on external computers is another amazing aspect. The researchers also demonstrated that ANYmal could overcome falls or slides in order to finish the obstacle course, in addition to being capable of handling a wide range of conditions.
Nevertheless, the research shows that robots are getting better at functioning in challenging real-world settings. That implies that they might soon be considerably more noticeable everywhere.
