Bipedal robot Cassie makes history by learning to run and covering 5 km
The Cassie robot, developed at Oregon State University (OSU), made history by covering a distance of 5 km in just over 53 minutes.
Cassie was developed under the guidance of robotics professor Jonathan Hurst over 16 months using a $1 million grant from the U.S. Defense Advanced Research Projects Agency (DARPA).
Since Cassie's debut in 2017, OSU students, funded by the National Science Foundation and DARPA's Machine Common Sense program, have worked with artificial intelligence professor Alan Fern to study the possibilities of machine learning for the robot.
Cassie became the first bipedal robot to use machine learning to control a running gait over open terrain, completing a 5 km run across the Oregon State University campus on a single battery charge and without a tether.
“Students at the Dynamic Robotics Laboratory in OSU's College of Engineering combined expertise in biomechanics and existing robot control approaches with new machine learning tools. This kind of holistic approach will make it possible to reach animal-like levels of performance. It is incredibly exciting”
The robot's knees bend, resembling an ostrich. It learned to run using a so-called deep reinforcement learning algorithm. Running requires dynamic balance, the ability to maintain equilibrium while changing position or otherwise being in motion. Cassie learned incredibly subtle adjustments to stay upright while moving.
“Cassie is a very efficient robot thanks to how it was designed and built, and we were really able to push the limits of the hardware and show what it can do”
Deep reinforcement learning is a powerful artificial intelligence method that develops skills such as running, jumping, and walking up and down stairs. One day, stably walking robots will become commonplace, like cars. The science and understanding of legged locomotion was the limiting factor, but Oregon State University's research has made several breakthroughs.
ATRIAS, developed in the Dynamic Robotics Laboratory, was the first robot to reproduce the dynamics of human walking. It was followed by Cassie, and then the humanoid robot Agility Digit.
Hurst said: "In the not-too-distant future, everyone will see and interact with robots in many places of their daily lives, robots that work alongside us and improve the quality of our lives." According to him, in addition to logistics work (for example, package delivery), bipedal robots will eventually have the intelligence and safety to help people in their own homes.
Cassie is often compared to an ostrich because of its gait and legs. The robot has three degrees of freedom in its hips and powerful, flexible ankles, but its knees bend in only one direction. All of this creates not only a natural gait but also a control method similar to our own. As with ATRIAS, the fast and careful placement of its feet makes Cassie resistant to slipping, tripping, and random malicious actions by graduate students.
The entire chassis is very lightweight and designed to absorb shocks naturally, as we do when walking. However, the robot is sturdy enough that it does not need safety harnesses in case of a fall. This is impressive for a robot built from scratch in a not very long period of time.
During the 5 km that Cassie ran in 53 minutes and 3 seconds, there was a reset that cost 6.5 minutes after two falls: one due to the computer overheating, the other because the robot was asked to turn at too high a speed.
Cassie itself serves as the foundation for Agility Robotics' Digit delivery robot. Oregon State University engineers were able to train Cassie in a simulator so that it could climb and descend stairs without using cameras or lidars.
However, bipedal devices are not an easy task. Their great advantage is that they can easily move through spaces designed for people, such as stairs and narrow corridors. But having two legs means they are inherently less stable, especially compared to quadrupedal robots such as those from Boston Dynamics.
The Cassie development team is confident this may change soon thanks to new deep learning methods that will allow bipedal robots to develop new stabilization techniques while moving. Who knows, maybe one day you will see such robots on your morning jog.
Such efficient and reliable bipedal robots that can go where people go will benefit countless industries, as well as dangerous work such as reconnaissance in disaster zones or inspection of nuclear power plants.
