(ORDO NEWS) — A group of researchers from the Universities of Illinois and Stanford have developed a training method that should improve the ability of mobile robots to navigate safely in crowded areas.
This method is based on the use of people around the robot as indicators of potential obstacles. The robot sees how people move and draws conclusions about hidden obstacles and the best route.
The fact that robots use humans as sensors for invisible obstacles sounds rather unexpected. But in principle, we must learn to help each other: robots – to us, we – to robots
A team of researchers from the University of Illinois at Urbana-Champaign and Stanford University have developed a deep reinforcement learning method that should improve the ability of mobile robots to navigate safely in crowded areas.
This method is based on the idea of using the people around the robot as indicators of potential obstacles.
Masha Itkina, co-author of the work, says : “We ourselves draw conclusions about the environment by observing the movement of people, thus considering people as sensors.
For example, if we see a driver brake hard, we can assume that a pedestrian has run into the road in front of that driver.”
Most of the previously developed models that consider people as sensors were created for use in urban environments to improve the safety of autonomous vehicles. The new model is designed to improve the ability of a mobile robot to navigate a crowd of people.
Crowd navigation tasks tend to be more difficult than urban driving tasks, since the movement of a person in a crowd is much more difficult to predict.
Optimal robot imperfection
Ye-Ji Moon, lead author of the study, says: “First, we imagine the robot’s environment as a map, it’s like a bird’s eye view. Occupied cells are marked on the map.
Thus, a system of obstacles around the robot is formed and the optimal route is laid.” During training, the robot tries to find this optimal route, focusing on moving and static obstacles, and builds its own map.
Unexpectedly, scientists found that the robot creates rather imperfect maps, on which the location of observed and covert agents is plotted rather inaccurately.
But these imperfect maps allow the robot to find routes that are close to optimal. It turned out that the robot evaluates only the location of the nearest “critical agents”, which are hidden and can block the robot’s path to the target.
“This result means that the most complete map is not necessarily the best for navigating in a partially observable, crowded environment. It turned out that it was enough to focus on a few potentially important agents,” Moon says.
“The main motivation for this work was to capture human intuition when navigating through crowds,” Itkina added. “We hope to better understand humans in order to improve the capabilities of robots.”
—
Online:
Contact us: [email protected]
Our Standards, Terms of Use: Standard Terms And Conditions.