Controlling a Drone After Sudden Rotor Failure #ICRA2022

Dr. Sihao Sun discusses his award-winning analysis within the space of controlling the flight of a drone when confronted with a sudden rotor failure.

Typical analysis on this space addressed the case the place one of many 4 rotors in a quadrotor out of the blue, spontaneously stops working. This earlier analysis doesn’t take into full account real-life eventualities the place rotor failure is widespread. This consists of collisions with different drones, partitions, birds, and working in degraded GPS environments.

Dr. Sihao Sun

Dr. Sihao Sun is a postdoctoral analysis assistant on the Robotics and Perception Group (RPG) in University of Zurich directed by Prof. Davide Scaramuzza. Currently, he’s engaged on management and notion for aerial robots (drones).

In December 2020, he acquired his PhD diploma in Aerospace Engineering from the Control and Simulation Group of Delft University of Technology. His works on quadrotor fault-tolerant flight management have been featured by respected media, comparable to IEEE Spectrum.

Links

transcript

Abate: [00:00:00] Can you inform me a bit of bit about your analysis?

Sihao Sun: I’ve a really broad analysis curiosity … the article of the analysis is areas of robotics like quad-copters, but in addition hybrid and even flapping drones.

Abate: What’s a hybrid drone?

Sihao Sun: A hybrid drone is like these drones with hybrid configuration, like with a wing and rotors… So in that sense, you possibly can leverage the benefits of each configurations. … you are able to do this vertical takeoff touchdown and in addition forest flight with extra aerodynamic effectively.

Abate: So you possibly can go a lot additional distance in addition to have the power to go up and down simply.

And so that you had a paper that was right here at ICRA “Upset recovery of control for quad rotors, subjected to complete rotor failure from large initial disturbance”. Could you speak about that paper?

Sihao Sun: Yes. Yes. I believe that’s an outdated paper. It was revealed again in 2020. So that work can be about this “motor failure” downside.

But so the background is that earlier than that paper, the prevailing works, making an attempt to deal with the motor failure downside of the quad-rotor, at all times assume that the failure occurs when a drone is nearing a hovering situation.

So when a drone is hovering and abruptly one motor fails, or that assumption that it’s, it’s in some way affordable. But it’s nonetheless a bit of bit removed from actuality. Because think about in actuality when the motor [fails], it might be due to some collision with different drones or with birds or one thing.

So that collision would change the angle of the drones dramatically.

Abate: So what you’re saying is that if the drone is simply hovering within the air, it’s unlikely that the motor would simply randomly fail in that place. It’s extra probably that it’s truly due to a collision, which is the reason for…

Sihao Sun: Yes.

So even when the motor fails, when it’s hovering, the software program nonetheless wants a while to detect the failure of that motor. So that point will roughly take about 100 milliseconds. So as a result of you recognize that in that point, throughout that interval, the drone itself, it doesn’t know that one motor fails. So it’s nonetheless utilizing its unique flight controller and its unique [00:03:00] flight algorithms.

And that the entire system has modified. Drones are very more likely to flip over in that interval. Which additionally implies that even when the video occurs across the hovering situation, the fault tolerance algorithm will take over when the drone is already tipping over.

So that’s why we wish to look into this state of affairs when a drone is sort of the wrong way up… Existing approaches, earlier than that paper was revealed, they assume that the drone is close to hovering and they’re going to use that assumption of their algorithms, which implies that these algorithms will fail in real-life eventualities. So we’ve got validated in our experiments that [it is] very onerous to get better the drones when a drone is sort of flipping over.

So that’s the initiative about that paper. And then we design an algorithm that is ready to get better the drone, not solely close to hovering but in addition when the drone is the wrong way up and [we are] capable of get better in that state of affairs. It’s very difficult, however that will show that it really works.

Abate: This is for a quadrotor the place solely one of many motors fails?

Also simply to know how steady a quadrotor could be with simply three motors and not using a fault tolerance algorithm… Will they simply instantly fall? Fly erratically? And ultimately, hit a flooring or a wall?

Sihao Sun: They simply behave randomly. And additionally, relying on which, algorithms are getting used to regulate the drone, as a result of there are numerous totally different algorithms. A PID or LQR, additionally with these management strategies, they’re cascaded.

So for instance, even when you’ve got an an identical controller to regulate the angle. The controller to regulate the place may additionally be totally different. And these variations of different controller as properly additionally results in totally different behaviors of the motor failures, it’s very onerous to say what’s going to occur, however it’s very probably that the drone will crash.

Abate: Yeah. And so what precisely did your analysis do? And a few of your analysis was introduced right here on the 2022 ICRA, appropriate?

Sihao Sun: I’ve two papers revealed in ICRA 2022. One paper can be making an attempt to deal with the the wrong way up downside, however it’s utilizing one other algorithm referred to as non-linear MPC. MPC is the nonlinear mannequin predictive management, which can be extensively used within the robotics group.

But no one has ever tried to make use of that on this motor failure situation. So [both of the papers I mentioned] attempt to remedy this upside-down restoration downside.

But within the older model of the paper, they use a cascaded construction. So principally the outer loop controllers… We at all times separate the controllers within the interior loop and outer loop controller as a result of the interior loop [controller] can react a lot faster. While, the outer loop, [gives] high-level instructions to the interior loop, however they react a lot, a lot slower.

So in these cascaded buildings, typically, the outer loop doesn’t know any constraints from the interior loop. So generally you simply give a really unreasonable command to the interior loop, which the interior loop can’t observe very properly. So that’s the issue of this cascaded construction.

So for that outdated paper we revealed two years in the past in ICRA, this downside occurs generally, which makes the upset restoration controller not 100% profitable. So that’s why we resolved to the nonlinear NPC strategies, which we publish on this ICRA as a result of that technique doesn’t use this cascaded construction.

The motor pace command, the bottom command is being calculated utilizing the entire state house of data. And so it doesn’t undergo from this cascading concern. And we present that this paper can behave a lot, significantly better than the earlier one.

Abate: And does this require any sort of particular sensors, with high-frequency inputs to make this work?

Sihao Sun: No. It’s the identical as different analysis, like utilizing, utilizing VICON or GPS as exterior sensors for state estimation and an IMU onboard for the physique fee estimates. So it’s all the identical with others. So we don’t want to make use of every other particular sensors, totally different from different analysis teams.

Abate: And is the place and orientation — that’s partially IMU — Is it additionally utilizing a imaginative and prescient system?

Sihao Sun: No. It’s not utilizing a imaginative and prescient system for that work. We additionally get a paper that was revealed in RAL final 12 months. And it was additionally introduced in, ICRA 2021.

It’s about utilizing these onboard imaginative and prescient sensors to unravel this motor failure downside. I’m additionally very honored that that paper was additionally nominated as one of the best paper of RAL previous 12 months. That was introduced yesterday. In that paper, we’re making an attempt to unravel this downside from one other perspective.

So it’s not about this get better the drone from this upside-down orientation, however making an attempt to make use of onboard sensors as an alternative of exterior sensors, as a result of there’s additionally a sensible state of affairs. So think about the motor failures in an space which was very near a tall constructing and the GPS sign is degraded.

So you can’t depend on these exterior positioning sensors for state estimation. So in that very same case, we’ve got to make use of the onboard digital camera. But the motor failure case for quad-copter could be very difficult for these onboard imaginative and prescient sensors as a result of the drone should (have) quick pace. So it will need to have, yaw spin that’s over 4 revolutions per second I might say.

It’s fairly quick for normal vision-based estimation algorithms due to the movement board, due to the ill-posed situation. So in that paper, we tried to unravel that downside.

To evaluate the 2 kinds of cameras: one is the usual body digital camera. So we fine-tune that. And the opposite is the event-based digital camera, which is a brand new sort of neuromorphic sensor.

It’s additionally a extremely popular analysis subject within the laptop imaginative and prescient group and the robotic group. That form of sensor doesn’t undergo from this movement blur. So we evaluate these two sorts of sensors on this motor failure state of affairs and we conclude that with the usual digital camera it’s nonetheless doable to regulate [the quadrotor] general, even when the motor fails.

But the event-based digital camera can do higher, particularly within the decrease mild situations. That was the conclusion of that paper.

Abate: Well decrease mild situations… and in addition I think about if that is open air and it’s spinning in a short time and out and in of the solar which may trigger some publicity points with the usual digital camera.

So one of many large advantages of the occasion digital camera is that it may possibly work throughout a bigger vary of brightness after which additionally the “frames”… it’s not likely working the identical manner with a “fixed frame rate”. So movement blur is far much less of a problem.

Sihao Sun: Exactly.

Abate: So, there are a few totally different strategies that you simply proposed. Is the final one that you simply’re speaking in regards to the one that you simply suppose is probably the most viable to be used in precise business?

Sihao Sun: Well, you recognize, for every paper, they’re the one addressing a single downside, however for business, it’s important to mix all of them collectively. So, what I might suggest is for industries to use the “model predictive control” strategies for the drones as a result of we’ve got proven that that technique could be very helpful in these motor failure situations in contrast with different cascaded buildings.

But you too can use the onboard digital camera that’s fine-tuned or use an event-based digital camera for that, despite the fact that occasion cameras are fairly costly in the meanwhile… however maybe sooner or later, you need to use that.

Abate: Perhaps sooner or later when the value goes down with mass manufacturing, however not but for many use circumstances.

transcript

tags: Algorithm Controls, c-Research-Innovation, cx-Aerial, cx-Research-Innovation, Flying, podcast, Research, Robotics expertise