TLDR: Researchers optimized the running gaits of the bipedal robot Cassie, comparing its mechanics to human running and finding surprising similarities. They then integrated these optimized gaits into a controller for a 100m dash, enabling Cassie to set a new Guinness World Record for the fastest 100m by a bipedal robot, demonstrating advanced sim-to-real reinforcement learning.
The world of robotics continues to push boundaries, and a recent breakthrough in bipedal locomotion has seen the robot Cassie achieve a remarkable feat: setting a new Guinness World Record for the Fastest 100m Dash by a Bipedal Robot. This achievement is detailed in a research paper titled “Optimizing Bipedal Locomotion for The 100m Dash With Comparison to Human Running” by Devin Crowley, Jeremy Dao, Helei Duan, Kevin Green, Jonathan Hurst, and Alan Fern.
Unlocking High-Speed Running for Bipedal Robots
For years, bipedal robots have struggled to achieve high-speed, efficient running comparable to humans or even quadrupeds. Previous attempts were often limited to slower speeds or required robots specifically designed for running, unlike general-purpose bipedal robots like Cassie. This research aimed to address how fast a robot like Cassie could run and how to develop a controller that maintains stability and efficiency across various speeds.
The first key contribution of this work was to systematically explore and optimize Cassie’s running gaits. Running gaits are essentially the robot’s pattern of movement, defined by parameters like ‘swing ratio’ (proportion of time a foot is in the air) and ‘stride frequency’ (number of footsteps per second). Unlike previous methods that used fixed or hand-tuned parameters, the researchers trained Cassie using reinforcement learning across a wide spectrum of these gait parameters and speeds. They then identified the most efficient combinations for different speeds. Interestingly, their findings suggested that higher speeds were best achieved with surprisingly lower stride frequencies, leading to longer, less frequent steps and a significant ‘aerial phase’ where both feet are off the ground. This contrasts with the intuitive idea that faster running means more rapid leg movements.
Comparing Robot and Human Running Mechanics
A fascinating aspect of this research was the comparison of Cassie’s optimized running mechanics to human running. Despite the obvious morphological differences between a robot and a human, the study found compelling similarities in their running behaviors. For instance, both Cassie and humans primarily increase their stride length to gain speed, with stride frequency remaining relatively flat or even decreasing at moderate speeds before increasing at top speeds. The ‘aerial time’ – the time spent with both feet off the ground – also showed strong agreement. While there were some differences, particularly in the effective ground reaction force at higher speeds (attributed to Cassie’s hip-roll motor limitations), the overall qualitative similarities suggest that the optimized gaits for Cassie are more ‘natural’ and efficient.
Also Read:
- New Method Boosts Robot Learning from Single Demonstration
- BaroPoser: Real-time Body Tracking Beyond Flat Surfaces with Everyday Devices
The 100m Dash Controller: A World Record Achievement
The third major contribution was integrating these optimized gaits into a full controller specifically designed for the 100m dash. This involved not just running at top speed, but also reliably starting from a standing position and coming to a complete stop after crossing the finish line, all without falling. The dash was broken down into five stages: standing start, acceleration, steady-state running, slowing down, and final standing stop.
To manage these different stages, the researchers used two specialized policies: a running policy for the main dash and a standing policy for the start and end. A critical challenge was smoothly transitioning between these policies. They developed precise timing mechanisms for these transitions. For example, when transitioning from standing to stepping, the controller waits for a specific moment in the gait cycle when the robot’s ground reaction forces are most similar to its standing behavior. For stopping, it waits until a foot is at its highest point (apex) to allow the standing policy to position the foot for a stable base.
These careful adjustments led to a remarkable success rate for the transitions. The culmination of this work was the official 100m dash attempts conducted in May 2022, where Cassie successfully completed the dash and established the Guinness World Record for Fastest 100m by a Bipedal Robot. While Cassie’s current top speed of over 4 meters per second is still about one-third that of top human runners, this research represents a significant leap forward in bipedal robotics, demonstrating the power of systematic gait optimization and sim-to-real reinforcement learning.
