Learning Aerobatics
Learning-based controller enables aerobatic maneuvers in insect-scale aerial robots
Learning Aerobatics
Learning-based controller enables aerobatic maneuvers in insect-scale aerial robots
Embedded Files
Learning-based controller enables aerobatic maneuvers in insect-scale aerial robots
Fast learning controller from MPC
Fast learning controller from MPC
To achieve aerobatic maneuvers, we first use a robust tube Model Predictive Controller (MPC) to generate flight demonstrations (in the simulation). A neural network then reproduces this behavior through imitation learning, aiming to preserve the effectiveness and disturbance tolerance of the original controller while being much faster to evaluate.
Extreme acceleration
Extreme acceleration
During aggressive turns, the robot reaches accelerations of 11.7 m/s², a 2.5-fold increase over previous insect-scale flight results. This allows the robot to rapidly accelerate, brake, and redirect its motion within a very small space.
High-speed saccade
High-speed saccade
The robot performs rapid lateral saccades at speeds of ~2 m/s, nearly 4.5× faster than previous results on the same platform. Despite the aggressive maneuver, the robot maintains accurate trajectory tracking throughout the flight.
Withstand wind disturbances
Withstand wind disturbances
The controller remains robust even under strong aerodynamic disturbances, which are not included in the demonstrations or the training stages. The robot can successfully execute aggressive saccades while exposed to wind speeds of up to 1.6 m/s without losing stable flight control.
Consecutive somersaults
Consecutive somersaults
The robot can repeatedly flip its body, completing 10 consecutive somersaults in only 11 seconds. This demonstrates a combination of agility, precision, and stability that was previously difficult to achieve at the insect scale.
High-fidelity simulation
High-fidelity simulation
We developed a high-fidelity simulation environment that closely reproduces the robot’s fast, nonlinear flight dynamics. The simulated motion matches the experiments well, making the simulator a useful platform for developing and training learning-based controllers before deployment on the real robot.
Speed & agility without sacrificing precision
Speed & agility without sacrificing precision
The robot can track trajectories far faster (4.5 times) and more aggressively (2.5 times) than previous insect-scale flyers while still maintaining precise control. Despite these much more demanding trajectories with body tilts approaching 50°, the robot maintains tracking errors of only a few centimeters, showing that substantially higher agility can be achieved without sacrificing precision.
Page updated
Google Sites
Report abuse