Fastest 100 m by a bipedal robot
Who
Oregon State University Dynamic Robotics Laboratory
What
24.73 second(s)
Where
United States (Corvallis)
When

The fastest 100 m by a bipedal robot is 24.73 seconds, achieved by Oregon State University Dynamic Robotics Laboratory (USA) in Corvallis, Oregon, USA, on 11 May 2022.


The bipedal robot Cassie, built by Agility Robotics, completed a 100 m dash from standing and returned to standing without falling. It was controlled by a neural-network trained for about a year in simulation compressed to one week in real time. The final controller was specialized for this task by Devin Crowley, who operated Cassie and orchestrated the record attempt. Stability and the sim-to-real transfer from simulation-based training to real-world deployment are among the Dynamic Robotics Laboratory (DRL)’s biggest challenges. The DRL has achieved very fast speeds on Cassie in simulation, but imperfect modelling leads to lesser performance on hardware.

There are other Guinness World Records titles for fast robots, but this one is based on a standard competitive human task. It is not just a measure of momentary top-speed – it is effectively average speed sustained across a certain distance and under the significant constraints that it must start in a standing pose and return to that pose after crossing the finish line. It cannot simply run 100 metres and crash.

Similar to the takeoff and landing of an airplane, starting and (especially) stopping were the most challenging hurdles of this endeavor. The neural network controller for the dash was specialized for running and unable to make Cassie stand still. To address this the DRL transitioned between two neural networks: one that can run, and another that can stand. Getting it to work gracefully was a matter of encoding proper timing for these transitions. Given the realistic nature of the task, establishing this record is a concrete milestone in robot locomotion and real-world capability.

This work is the culmination of years in bipedal locomotion research by the Dynamic Robotics Laboratory at Oregon State University with contributing members past and present including Devin Crowley, Jonah Siekmann, Yesh Godse, Jeremy Dao, Helei Duan, Kevin Green, Ryan Batke, Fangzhou Yu, Aseem Saxena, Ashish Malik, and Mohitvishnu Gadde, together with lab co-directors Alan Fern and Jonathan Hurst.