New research teaches robots to overcome obstacles

Most four-legged robots are trained to regain their balance if they stumble upon an obstacle. In an effort to develop a cleaning robot, Vietnamese-born PhD student Joanne Truong at the School of Interactive Computing at the Georgia Institute of Technology (GIT) and her two colleagues Naoki Yokoyama and Simar Kareer are training their robot to step over cluttered objects it might encounter in the home, Tech Xplore recently reported.

(From left) Naoki Yokoyama, Joanne Truong and Simar Kareer working with the four-legged robot

According to the research team, four-legged robots equipped with "blind" movement controllers tend to react more to avoid falling when they step on an object.

Meanwhile, the research team applied a new approach, providing live images for the robot to step over obstacles, by combining the navigation policy with the image-based locomotion policy. This approach helped the robot step over obstacles in a simulated cluttered environment with a success rate of up to 72.6%.

The robot can learn on its own and does not imitate any pre-existing behavioral patterns. The researchers say the model is scalable and can be applied immediately without much fine-tuning. The policies instruct the robot to avoid objects as it moves from one place to another and to use its legs to step over objects, including how to lift its legs to the appropriate height.

'Robot dog' overcomes long, bumpy roads without falling

According to the team, conventional quadrupedal robots can only see the real world through a front-facing camera and cannot see objects near their feet. The team incorporated memory and spatial awareness into the network to teach the robot exactly when and where to step over obstacles. If the object was too high, the robot could go around it. “We found that this method navigates very well, and even if the robot goes the wrong way, it knows it can back up and return to its original position,” Truong said. The team also taught the robot which objects it should step over, such as toys, and which objects it should go around, such as tables and chairs.

The team's findings could also help robots navigate real-world outdoor environments, choosing paths based on their owners' wishes to avoid muddy or rocky terrain.

The research won first prize at a robotics workshop at the Robotics Learning Conference 2022 in New Zealand. The research will be presented at the IEEE International Conference on Robotics and Automation in London, UK, from May 29 to June 2.

Source link: https://thanhnien.vn/nghien-cuu-moi-day-robot-vuot-chuong-ngai-vat-185230527204029629.htm