In the future, small flying robots could be deployed to help search for survivors trapped under rubble after devastating earthquakes. Like real insects, these robots can fly through tight spaces that larger robots cannot reach, while simultaneously avoiding stationary obstacles and falling debris.
So far, aerial microrobots have only been able to fly slowly along smooth trajectories, a far cry from the quick and agile flight of real insects.
MIT researchers have demonstrated an aerial microrobot that can fly with speed and agility comparable to its biological counterpart. The joint team designed a new AI-based controller for the robot bug, allowing it to follow gymnastic flight paths, including performing continuous body flips.
The two-part control scheme, which combines high performance and computational efficiency, increased the robot’s velocity and acceleration by about 450 percent and 250 percent, respectively, compared to the researchers’ best previous demonstrations.
The fast robot was agile enough to complete 10 consecutive somersaults within 11 seconds, even when wind turbulence threatened to throw it off course.
Credit: Provided by Soft and Micro Robotics Institute
“We want to be able to use these robots in scenarios where traditional quadcopter robots are difficult to fly, but insects can navigate. Now, with a biologically inspired control framework, the flight performance of our robots can be improved by speed. “This is a very exciting step toward future goals,” said Kevin Chen, associate professor in the Department of Electrical Engineering and Computer Science (EECS) and head of the Department of Soft and Microrobotics. Laboratory within the Research Institute for Electronics (RLE) and co-senior author of the robotics paper.
Chen is joined on the paper by co-lead author Yi-Hsuan Hsiao, an EECS MIT graduate student. Andrea Tagliabue PhD ’24; Owen Matteson, graduate student in the Department of Aerospace Engineering (AeroAstro); EECS graduate student Suhan Kim agrees. Tong Zhao Meng ’23; co-senior author Jonathan P. Howe is a professor of engineering in the Ford School of Aerospace Engineering and a principal investigator at the Institute for Information and Decision Systems (LIDS). This research today scientific progress.
AI controller
Chen’s group has been working on developing insect robots for more than five years.
They recently developed a more durable version of the small robot, a microcassette-sized device that is lighter than a paperclip. The new version features larger flapping wings that allow for more agile movements. They are powered by a series of fluffy artificial muscles that allow them to flap their wings at extremely high speeds.
However, the controller, the robot’s “brain” that determines its position and tells it where to fly, is manually adjusted by humans, limiting the robot’s performance.
For the robot to fly quickly and aggressively like a real insect, it needed a more robust controller that could account for uncertainties and perform complex optimizations quickly.
Such a controller would be too computationally intensive to deploy in real time, especially given the complex aerodynamics of lightweight robots.
To overcome this challenge, Chen’s group collaborated with How’s team to co-create a two-stage, AI-driven control scheme that provides the robustness needed for complex, rapid operations and the computational efficiency needed for real-time deployment.
“As controllers have advanced with advances in hardware, we’ve been able to do more on the software side, but at the same time, as controllers have been developed, we’ve also been able to do more with hardware. As Kevin’s team demonstrated new capabilities, we demonstrated that we could take advantage of them,” Howe says.
As a first step, the team built what is known as a model predictive controller. This type of powerful controller uses a dynamic mathematical model to predict the robot’s behavior and plan the optimal course of action to safely follow the trajectory.
Although computationally intensive, you can plan difficult maneuvers such as mid-air somersaults, quick turns, and aggressive vehicle leans. This high-performance planner is also designed to take into account the force and torque constraints that the robot can apply, which is essential to avoid collisions.
For example, to perform multiple flips in succession, the robot must be slowed down so that the initial state is exactly appropriate to perform the flip again.
“If a small error creeps in and you try to do 10 flips with that small error, the robot will just crash. You need robust flight controls,” Howe says.
They use this specialized planner to train “policies” based on deep learning models to control the robot in real time, through a process called imitation learning. Policies are the robot’s decision-making engine, telling it where and how to fly.
Essentially, the imitation learning process compresses a powerful controller into a computationally efficient AI model that can run very quickly.
The key was to have a smart way to create enough training data to teach policy everything needed for offensive operations.
“A robust training method is the secret to this technique,” Howe explains.
AI-driven policies take the robot’s position as input and output control commands such as thrust and torque in real time.
insect-like performance
In their experiments, this two-step approach allowed the insect-scale robot to fly 447 percent faster while increasing its acceleration by 255 percent. The robot was able to complete 10 somersaults in 11 seconds, and the tiny robot never strayed more than 4 to 5 centimeters from its planned trajectory.
“This study shows that traditionally speed-limited soft robots and microrobots can leverage advanced control algorithms to achieve agility approaching that of natural insects and large robots, opening up new opportunities for multimodal locomotion,” Xiao said.
The researchers were also able to demonstrate saccadic movements, which occur when an insect pitches very aggressively, rapidly flying to a certain position, then pitching in the opposite direction and coming to a stop. This rapid acceleration and deceleration helps the insect to locate itself and see clearly.
“This biomimetic flight behavior could be useful in the future when we start equipping robots with cameras and sensors,” Chen said.
Adding sensors and cameras to allow microrobots to fly outdoors without being connected to complex motion capture systems will be a major area of future work.
The researchers also hope to study how onboard sensors can help robots avoid collisions with each other and adjust navigation.
“For the micro-robot community, we hope this paper signals a paradigm shift by showing that it is possible to develop new control architectures that are both high-performance and efficient at the same time,” says Chen.
“This work is particularly impressive because these robots still perform precise flips and high-speed rotations despite large uncertainties caused by relatively large manufacturing tolerances in small-scale manufacturing, wind gusts exceeding 1 meter per second, and even power tethers wrapping around the robots when performing repeated flips,” says Sarah Bergbreiter, a professor of mechanical engineering at Carnegie Mellon University who was not involved in the study.
“Although the controller currently runs on an external computer rather than on-board the robot, the authors demonstrate that similar but less precise control policies may be feasible even with the more limited computational power available to insect-scale robots. This is interesting because it suggests future insect-scale robots with agility approaching that of their biological counterparts,” she added.
This research was funded in part by the National Science Foundation (NSF), Office of Naval Research, Air Force Office of Scientific Research, MathWorks, and the Zakharchenko Fellowship.
