A self-taught AI robot can already bunny hop a bike better than we do

Watch as this self-taught AI bot rides a balance bike, bunny hops, wheelies, and rides backward, thanks to reinforcement learning.

Learning how to bunny hop a bike takes time. That goes for humans… and robots, too.

The Robotics and AI (RAI) Institute recently shared a video of a robot attached to a Specialized Hotwalk balance bike, which they’re calling the Ultra Mobility Vehicle (UMV). Watch as the UMV manages to not only balance, but also hop up onto a table, do a wheelie, and ride backward, all without human intervention.

The crazy thing is that, according to this article, the UMV doesn’t utilize a gyroscope or a stabilization system for balance. The bike can only move forward and backward and turn the front wheel. In theory, it seems like this might put the bot at a disadvantage over a human rider who can also shift weight to maintain balance, in addition to turning the wheel and propelling the bike.

Screenshot.

The RAI UMV demonstrates reinforcement learning, a machine learning technique where a bot learns to make decisions by interacting with an environment, receiving feedback in the form of rewards or penalties for its actions, and gradually optimizing its behavior to maximize rewards over time. This is basically how humans learn to do new things, too.

The article doesn’t specifically address how the UMV was trained or how long the training took, but it likely involved sensors to detect if the bike falls over or hits the table and running hundreds or thousands of trials until the UMV could complete the course without falling. On the reward side, the UMV was likely programmed to hit certain targets like following the orange line or jumping above a certain height while moving forward at a certain speed.

If you think about it, this is basically how we all learned to ride a bike. Falling over hurts (penalty), so we avoid it to maximize our reward (fun). Then, we rinse and repeat to learn new tricks.

According to the RAI Institute team, the most difficult move for the UMV was balancing while riding backward. While the UMV lab performance seen in the video is impressive, it sounds like further work is needed for the bot to ride off road.

If I had to guess, it’s only a matter of time before the UMV is ready for Redbull Rampage.