New kernel learning method trains quadrotors to adapt in 50 seconds

Researchers have developed a new algorithm for domain-adaptive policy learning that lets quadrotors handle non-stationary disturbances like sudden wind gusts, payload shifts, and ground effects. The method models unknown disturbances using a differentiable kernel approximation based on random Fourier features. During offline training, kernel coefficients and bandwidths are randomly sampled to generate diverse disturbance profiles, and the control policy is optimized via differentiable simulation with analytical gradients. Remarkably, training takes only 50 seconds on an RTX 4090 GPU. During deployment, the policy adapts in real time by updating kernel parameters through online least-squares estimation, enabling the drone to handle changing conditions without retraining.

In experiments, the algorithm was tested on Crazyflie quadrotors performing trajectory tracking tasks under high-fidelity simulations and real hardware tests. It successfully managed complex aerodynamic effects, wind, ground effects, and payload fluctuations, outperforming traditional adaptive control methods. This work bridges the gap between offline training efficiency and online adaptation speed, making it practical for real-world drone applications where environmental conditions are unpredictable. The approach could extend to other robotics domains requiring rapid adaptation, such as autonomous vehicles or manipulators facing variable loads and conditions.