Bracing Humanoid Robots for the Inevitable—Falling

After having worked for about 20 hours with a 99% success rate at ProMat, Digit, a humanoid robot by Agility Robotics, fell. And it just looks like Digit goes, "I'm tired of doing this job," and just falls down, and people would be like, "I totally feel you, man." When, in reality, it really is okay and, in fact, important that these humanoid robots fall.

It is part of the learning process, just as it is for us humans. We learn to walk and balance ourselves by falling and getting back again, and just like that, robots can learn in a similar way. It allows them to learn about their environments, test their limits, and teach themselves to solve their own problems, which in this case would be picking themselves up.

When a robot falls, it can examine what went wrong and then try to adjust its movements to avoid similar mistakes later on. And this helps them improve their balance and coordination.

Designing Robots to Get Back Up

The idea is to not need human intervention for these processes to continue, and then failure happens. In this case, a robot falling and picking itself back up is very useful for maintaining productivity, especially in factories. And so some of these humanoid robot makers are designing their robots to get back on their feet from a prone position.

Just after Boston Dynamics retired their hydraulic humanoid robot Atlas, they released an electric humanoid robot with the same name, Atlas. In their demo video, they showcased the robot lying face-down on the floor, then standing up right on its own.

It may seem uncanny the way Atlas does it—standing on its own—with how very different from the way humans would do it. The idea in this case is to minimise the area of space the robot utilises to get up, as not all environments are structured to provide so much area.

Agility Robotics, on the other hand, employs the conventional human-like method to get up, with the arms pushing the body back up. What's interesting, however, is how it first tries to put itself in a familiar position to then proceed to get up.

The company has been using reinforcement learning to help fallen robots right themselves. Agility shut off Digit’s obstacle avoidance in the above video to force a fall. In the video, the robot uses its arms to mitigate the fall as much as possible. It then utilises its reinforcement learnings to return to a familiar position from which it is capable of standing again with a robotic pushup. _Source

The Importance of Transparency

A lot of work goes into the demo videos of these humanoid robots we see around. Sometimes it takes a week, perhaps years, for some to achieve the perfection they show us. In those times, a lot of falling happens. And then, the value of how much effort goes into building these robots is sometimes undermined.

A robot that has fallen many times and learned how to pick itself back up is likely to fall less in the future. And so, seeing and knowing how much goes into actually building and training these robots can help people appreciate them a little more, rather than assuming that it's easier than it actually looks.

For that reason, we see many videos of Boston Dynamics' robots failing as well as their successes, which is really good as it provides more insight into their work. The same goes for Agility Robotics, with their videos all over social media.

Falling is Part of Learning

As long as these robots learn what to do when they fall, they will only get better and probably fail a whole lot less than humans do, ultimately turning into the machines we want them to be—highly effective and efficient.

I imagine these robotics companies putting their humanoid robots in environments to enable them to fall a lot more and learn. With time, high success rates will be achievable when they are fixed into workflows.