This article is part of our exclusive IEEE Journal Watch series in partnership with IEEE Xplore.
Four-legged robots are making great strides in their ability to safely traverse complex terrains. In a recent advance, researchers in Hong Kong have developed a novel mapping model for quadrupedal robots that allows them to autonomously crawl under and leap over significant obstacles in order to arrive at its desired endpoint.
The researchers describe the robot, which uses multilayered mapping to understand its environment, in a study published 4 August in IEEE Robotics and Automation Letters.
Agile two- and four-legged animals can adapt to diverse terrains. Robots that can traverse similarly complex environments are appealing because they could complete missions that would be dangerous for humans to do, like monitoring and assessing unstable rubble sites after an earthquake. But ensuring that robots can effectively map out complex environments, as well as handle the various types of obstacles (leaping across gaps and or scaling high objects, for example) is challenging.
Advanced Terrain Mapping for Robots
Peng Lu, an assistant professor at the University of Hong Kong, along with his postdoctoral student Yeke Chen and other team members sought to create a robot capable of overcoming these hurdles. To help their robot perceive its surroundings in detail, they developed a model that creates a multilayer elevation map from the robot’s sensor data. The map can capture the characteristics of a wide range of terrains using lidar data.
The team used simulations to train the robot to recognize different terrains it may encounter in the real world. This includes very challenging terrains to navigate, such as a gap that it must jump across or crawling under obstacles with an overhang jutting out. If the robot has missing sensor data, it can compensate to some extent with estimations, based on its training data.
“Through learning different skills in simulation and knowledge distillation, the robot is able to switch among different skills to traverse through different obstacles,” says Lu.
In their study, the researchers tested their mapping technique using a Unitree Go1 robot in a series of indoor and outdoor experiments, where it had to autonomously crawl, climb, or jump to overcome obstacles.
“The results show that the multilayer elevation map can effectively represent various complex terrains, which allow a robot to easily understand the environment,” says Lu, noting the robot also succeeded in autonomously switching between modes—crawling, jumping, and climbing—as needed.
He adds that the robot inadvertently has path-planning abilities, even though it was not programmed to have them. For example, when the robot encounters obstacles that are too high to pass, it moves around the obstacle, and therefore finds a path through the environment on its own through trial and error.
Lu notes that while a key strength of this robot is its ability to navigate diverse and difficult terrain, it can rely only on the data that it has already been trained on, and cannot learn directly from real-world data.
Lu says his team may commercialize the robot for inspection scenarios, such as construction sites, and plans on using real-world data to further enhance the robot’s ability to cope with any type of terrain.
