Machine-learning robot athlete sets record time for 5km run

'Cassie' achieved the feat in just over 53 minutes, which may be unimpressive in human terms - particularly during the Tokyo Olympics 2020 - but it is significantly faster than other ML bipedal robots.

Cassie was developed under the direction of robotics professor Jonathan Hurst with a 16-month, $1m (£700K) grant from the Defense Advanced Research Projects Agency (DARPA).

It was the first bipedal robot to use machine learning to control a running gait on outdoor terrain and completed the 5K run on Oregon State’s campus untethered and powered by a single battery charge.

“The Dynamic Robotics Laboratory students in the OSU College of Engineering combined expertise from biomechanics and existing robot control approaches with new machine learning tools,” said Hurst. “This type of holistic approach will enable animal-like levels of performance. It’s incredibly exciting.”

Cassie, with knees that bend like an ostrich’s, taught itself to run with what’s known as a deep reinforcement learning algorithm.

Running requires dynamic balancing - the ability to maintain balance while switching positions or otherwise being in motion - and Cassie has learned to make infinite subtle adjustments to stay upright while moving.

“Cassie is a very efficient robot because of how it has been designed and built and we were really able to reach the limits of the hardware and show what it can do,” said Jeremy Dao, a PhD student who worked on the project.

“Deep reinforcement learning is a powerful method in AI that opens up skills like running, skipping and walking up and down stairs,” added Yesh Godse, an undergraduate in the lab.

Hurst said walking robots will one day be a common sight, adding that the limiting factor up to now has always been the science and understanding of legged locomotion.

“In the not-very-distant future, everyone will see and interact with robots in many places in their everyday lives, robots that work alongside us and improve our quality of life,” he said.

In addition to logistics work such as package delivery, bipedal robots eventually will have the intelligence and safety capabilities to help people in their own homes, Hurst believes.

During the 5K, Cassie’s total time of 53 minutes, 3 seconds included about 6½ minutes of resets following two falls: one because of an overheated computer, the other because the robot was asked to execute a turn at too high a speed.