You may have seen this video go viral this week. It’s exciting to watch and marvel at. So, why is this inverted pendulum thing so challenging? Go find a broom and balance it on your chin. Notice how often you have to move to keep the broom stable above you. Think about how challenging it would be to, not only catch that broom on your chin, but to continue to balance it once it lands. In order to accomplish that with a friend, they’d have to throw it just right and you’d have to constantly correct your position in an attempt to keep it stable upon landing.
An Institute for Dynamics Systems and Control multirotor catches an inverted pendulum. Perfectly.
I’ve thoroughly enjoyed watching multirotor technology change over the past few years. Brushless motors are being built with more powerful rare-earth magnets, leading to much higher thrust-to-weight ratios, and the components required to keep a multirotor stable–accelerometers and gyroscopes–have gotten smaller and cheaper. It’s now easier than ever for hobbyists to get into the air, and for researchers to accomplish incredible tasks like we see here.
We’ve covered a few “wow” projects in the past, like the synchronized swarm from Kmel Robotics and UPenn’s GRASP Lab. While the fine folks in Pennsylvania have focused on multirotor trajectory planning and movement, the Institute for Dynamics Systems and Control at ETH Zurich has been exploring multirotor interactions with inanimate objects. In the past, they’ve had quadrotors juggling a ball, assembling structures, and balancing an inverted pendulum. Now an ETH Zurich student has taken their work with pendulums and multirotors to a new level.
The heart of the challenging maneuver seen in this video is feedback control, more specifically, PID control loops. As soon as the first quad launches the pole into the air, the vision system tracks it and predicts the impact position–where the “catching” quad needs to end up. As mentioned in the video, the “throwing” quad angles the pole so that it will rotate into an upright equilibrium on its own once it hits the landing pad, so it’s companion has to do minimal work to keep it balanced. Of course, the planned throwing angle isn’t enough to guarantee success, so the trajectory of the pole is calculated 50 times a second, allowing the other quad to get itself into the right position in time to make the catch.
By recording and analyzing the movement of the pole throughout its journey and applying learning algorithms, the system actually improves over time. To me, this is the most exciting part of the research. Think about where machine learning can be applied elsewhere. There are a lot of things that robots can, and will, do better than humans.
I’d like to hear what you think about all this: Are there any real-life applications for balancing an inverted pendulum on a multirotor? I see the Flying Machine Arena (FMA) work as more of a stepping stone to better multirotor control and interactions between robots and their surroundings. When will multirotors be able to accomplish tasks like this without the aid of external vision systems? How long will it be before they’re playing catch with a double inverted pendulum? Let me know what you think in the comments below.
Want more technical information? You can find the FMA’s publications here and KMel Robotics Co-President Daniel Mellinger’s publications here. For more information on PID Control of an inverted pendulum, check out this paper.