Researchers at the Carnegie Mellon Robotics Institute have introduced a learning-based control framework called Agile But Safe (ABS). The framework– developed and programmed by Tairan He, Chong Zhang, Wenli Xiao, Guanqi He, Changliu Liu, Guanya Shi– enables quadrupedal robots to move in a collision-free manner in confined indoor and outdoor environments.
When programmed with ABS, the Unitree Go1 quadruped robot can avoid both dynamic and stationary obstacles while moving at high speeds, demonstrating crucial sensing and computational abilities honed by the research team at CMU. The students tested its capabilities through a series of physical tests, including having humans jump in front of it and placing obstacles such as garbage cans and strollers in its path to test the reaction time and mobility features. Each physical test was designed to mimic real-world situations that the robot could encounter, such as someone opening a door suddenly.
“Our robot can easily navigate through super confined spaces like offices, narrow corridors, and in the snow and also on the grass,” said Tairan He, PhD student at the Robotics Institute. “It can go through a lot of static and dynamic obstacles like humans and suddenly appearing obstacles like boxes.”
Traditionally, agile robots have been programmed to move at slower speeds to reduce safety concerns. However, a key component to the teams’ research was to maintain the robot’s speed and agility while not compromising its ability to move through environments safely.
“Agile policy has a basic capability of collision avoidance,” said He. “But the main contribution of ours into the agile policy is that it can run fast. Up to 3.1 meters per second.”
Agile policy alone cannot guarantee that a robot will move about an environment safely. To aid with this the CMU team created a recovery policy. The computer vision maps the front view of the robot and constantly checks the distance between the robot and an obstacle to ensure safety. Once the robot predicts something will be in the way, it goes into the proper mode of recovery such as dodging, adjusting speed, or lateral movements to avoid the obstacle.
“Since the robot has a very clear vision of what will happen in the future, they can take actions before the future actually happens,” said Changliu Liu, assistant professor of robotics at CMU. “That’s resulted in the robot being much more fluent in collision avoidance. They do not need to stop and think. They can think on the fly.”
The team’s project was accepted into the 2024 Robotics: Science and Systems (RSS) Conference. RSS has an expansive history of uniting researchers from all disciplines of robotics and showcases some of the best work occurring in the robotics world.
Read the team’s paper here and learn more about Agile But Safe from the featured video.
For More Information: Aaron Aupperlee | 412-268-9068 | aaupperlee@cmu.edu