In the field of aerial microrobots, the challenge of speed and agility has long remained a dream. Although natural flyers like bees, flies, and dragonflies can execute extremely swift aerobatic maneuvers as if they were no longer than a mere step in their aerodynamic dance, their robotic kin have thus far remained rather grounded in terms of speed and flexibility of control algorithms adequate to overcome turbulent flows in small spaces. However, a combination of the MIT Soft and Micro Robotics Laboratory and the Laboratory for Information and Decision Systems has marked a decisive step in the direction of creating a small-scale robot capable of a high-speed aerobatic flight.
1. Two-Step AI Control for Real-Time Agility
The innovation of this team’s approach is to couple in a two-stage control system the strength of model predictive control (MPC) with the speed of deep learning algorithms. In the first step of their approach, a nonlinear robust tube MPC develops an optimal trajectory for an aircraft subject to force and torque, even when the aerodynamic effects are uncertain. The “expert” trajectory planner has the capability to coordinate complex actions such as ten straight somersaults in a row, saccades, and aggressive turns involving both pitch and roll—though it takes too long to run in real time for an aircraft to directly follow it. In the second step of their method, imitation learning reduces an MPC policy to a microsecond neural network controller that provides force and torque control directly. According to Jonathan P. How, “The robust training method is the secret sauce of this technique.”
2. Soft Actuators and High-Frequency Flapping
The agility of the robot is achieved through soft robotics technology. The four separate flapping-wing units, in combination with dielectric elastomer actuators, flap at a rate of 330 Hz, which is 70% faster than in fruit flies. The artificial muscles react in only 3 ms to the control commands, and this is essential for insect-like movement. The bioinspired design of the propulsion system replicates the power density and high-speed kinematics found in nature, enabling the robot to achieve extreme pitch and roll degrees.
3. Bioinspired Maneuvers: Saccades and Somersaults
In trials, the robot exhibited saccade motion reminiscent of fly strategies for stabilizing vision in cluttered spaces: rapid change of pitch to accelerate towards its target, followed by an inversion of pitch to decelerate. In addition, it executed, in succession, 10 somersaults in 11 seconds while keeping the deviation of its trajectory within 4-5 cm. These accomplishments involved accurate torques close to actuator saturation, while also being subject to tether Quinn §ling of up to 1 m/sec.
4. Robustness in Turbulent Conditions
Based on knowledge of active and passive stabilization in insect flight, the controller was able to deal with disturbances of magnitude equivalent to 160 cm/s of wind, more than 260% of previous disturbance rejection records. “The robust tube MPC is able to compensate in real time for some effects not modeled, such as downwash in a fast descent or tension in a flip due to a power tether, in order to keep the robot in a safe ‘tube’ of states.”
5. Precision Tracking of Complex Trajectories
The system performed well on complex routes: X-pattern flights with nine sharp turns in 5.5 seconds, figure eight routes with a velocity of 197 cm/s (446% improvement over subgram-scale robots), and circular sprints with 152 cm/s velocity. The RMS errors in position were below 3 cm, indicating that precision does not necessarily compromise with agility.
6. Integration with Autonomous Navigation Research
The control approach is consistent with recent work in vision-based UAV control using differentiable physics, where minimalist neural policies run on low-cost hardware. Although the current microrobot uses motion capture, subsequent versions will include sensors and cameras, allowing for unconstrained flight in GPS-denied outdoor environments, as well as multirobot navigation without collisions.
7. Consequence for Search-and-Rescue and Environment Observation
With insect-like agility, such robots could enter into areas where buildings are collapsed, or into heavier growths of vegetation, or other areas that currently are not accessible, to do reconnaissance work or place sensors in key positions by reconnoitering or delivering the sensor into a region that cannot be entered by larger robots or currently by living pollinators. This accomplishment combines the best of bio-inspired hardware, advanced control theory, and the power of learning-based compression. With their work, the MIT team has filled the gap between the raw computing power of MPC and the requirements imposed by real-time processing in insect flight and established a new benchmark: a soft microrobotic system that performs better than its natural model in certain areas.