four-legged robot ANYmal can overcome almost any obstacle
ANYmal – a robot on four “legs”, designed in the shape of a dog. It was created by a team of scientists from the Swiss higher technical school ETH with the aim of searching and rescuing people under rubble, for example, after natural disasters. Engineers improved the robot and taught it basic freerunning movements, one of the areas of parkour. This is an active sport that involves moving quickly through difficult obstacles such as walls, railings, building ruins and other obstacles. But this is a sport for people, but what can a robotic system do?
What about parkour and robot?
People who practice parkour have special dexterity and precision of movements, and they have excellent coordination. All this allows you to perform complex tricks. And although no robot has such abilities, ANYmal has learned to overcome obstacles, climb up and perform maneuvers in difficult conditions. All this is not just like that – the robot can work in places with unstable, rugged terrain or in destroyed buildings, both inside and outside.
ETH development team in 2019 presented an innovative reinforcement learning method that allows the robot learns based on experience gained and interaction with the environment. Last year, engineers improved ANYmal, making it more sensitive to its surroundings. In their experiments, the team tested three customized robots on training grounds simulating the surfaces of the Moon and Mars. In the future, designers from ETH plan to use ANYmal in space exploration missions, where such robots will be able to perform tasks to study territories or assist rovers. Each robot is equipped with a lidar sensor, which helps to navigate well in space.
The main purpose of the Scout is to view the environment using cameras and create a map using color filters. The Scientist model is equipped with an arm on which a MIRA analyzer (Metrohm Instant Raman Analyzer) and a MICRO microscopic meter are installed. MIRA can identify chemicals on a surface by light scattering, and MICRO can image these substances up close. The robot model, called Hybrid, is versatile and helps Scout and Scientist measure objects such as boulders and craters.
Despite significant progress in the development of ANYmal robots, they are still far from human levels of flexibility and dexterity. Nikita Rudin, PhD student at ETH Zurich and co-founder of the project, notes, that some researchers believe the limits of the physical capabilities of four-legged robots have been reached. “I have a different opinion. I firmly believe that the mechanics of such robots can be further developed,” assures Rudin, who himself has parkour experience.
Parkour-style movements are challenging for mechs, as they require not only dynamic maneuvers, but also motion control coupled with the ability to quickly adapt to a changing environment. To achieve success, ANYmal will have to be able to analyze circumstances, choose the optimal route and sequence of movements to maintain balance. And all this at the same time! To do this, the robot needs to use all programming skills in real time, resorting to limited computing resources.
To solve the ANYmal robot control problem, the ETH team developed an approach that combines machine learning methods with model control. It consists of three main components:
Perception module: This module processes data from cameras and the lidar device to assess the surrounding terrain and obstacles.
Movement module: This provides a pre-designed catalog of motions that allows the robot to overcome a variety of terrain types and obstacles.
Navigation module: This module is responsible for selecting the appropriate skills and movements to overcome specific obstacles and terrain. It uses intermediate commands to direct the movement module.
Nikita Rudin used machine learning to enable the robot to successfully jump and climb. These skills were developed through a series of trials and errors, as well as data obtained from the robot's camera and neural network training. Another researcher, Fabian Genelt, used a combination of machine learning and model control techniques to enable ANYmal to recognize and navigate obstacles such as cracks and openings in piles of debris. This helps the robot develop flexibility by using certain movements in unexpected situations.
During testing, the robot managed to learn how to descend from a box, which was 1 meter high, and then climb onto it again. ANYmal can also crouch to fit into narrow gaps and change its gait to maintain balance and prevent collapses. The team of engineers also tested the robot’s ability to walk on stairs, slopes and other uneven surfaces.
Not without difficulties
Despite the research team's successes, ANYmal's abilities remain quite limited when it comes to moving in real-world environments, such as parkour ranges or the rubble of destroyed buildings. The authors of the project note that they will have to test the robots in various random scenarios. The question of how ANYmal will manifest itself in more complex conditions remains open. To prepare for new tests, engineers need to configure and train an additional eight neural networks. The task is complicated by the fact that some of them are interconnected, so setting up one network entails changes in the others.
But ANYmal is already partially capable of working where it is necessary to make jumps, overcome obstacles and choose the optimal path to the target point. Members of the project team emphasize that the desire to achieve human-like dexterity helps to gain a deeper understanding of the limitations and capabilities of each component of the robot, as well as how ANYmal perceives its environment, acts and makes decisions.
