Tech

Watch this sightless bi-pedal robot walk down stairs flawlessly

It’s amazing what some simulated training can do.

For robots, bi-pedal locomotion is tough business. Now, imagine adding some stairs into the mix and taking away the ability to actually see where they’re walking.

🎥: IEEE Spectrum

That’s exactly the challenge researchers at Oregon State University set up for their robot, Cassie, made by Agility Robotics. And rather than having us tell you how Cassie performed, just watch this video of it tackling this formidable test for yourself.

Oregon State University

Cassie’s trial-by-stairs wasn’t just for fun, it was part of an ongoing effort by researchers to expand robots’ ability to use other senses for locomotion beyond just computer vision (which can be easily obstructed).

Equally impressive as Cassie’s ability to traverse stairs by feel and memory alone is the method by which researchers trained the robot.

Instead of having Cassie undergo a potentially dangerous crash course via actually climbing stairs, they used a method called sim-to-real Reinforcement Learning, which trained the bot virtually, allowing it to make mistakes without taking any IRL spills.

After the virtual training, Cassie was unleashed on sets of stairs it had never traversed before (10 trials ascending and 10 trials descending) and sight be damned, the bot faired pretty well.

100%

Cassie had a perfect success rate on descending stairs.

80%

Cassie faired slightly worse while ascending stairs.

Researchers say there’s still room for growth in Cassie’s performance. For instance, the bot could only walk at standard speed without faltering... and it isn’t particularly good at improvising.

In the future, however, researchers say it can be paired with computer vision to find out how much (or if) its ability to climb stairs improves, though it’s hard to imagine giving it the ability to actually see where one is stepping could hurt locomotion.

That means more of this.

And less of this.