Avirup Das*¹, Rishabh Yadav*¹, Sihao Sun², Mingfei Sun¹, Samuel Kaski1 ³, Wei Pan¹
¹ Department of Computer Science, The University of Manchester
² Department of Cognitive Robotics, Delft University of Technology
³ Department of Computer Science, Aalto University
* Equal Contribution
DroneDiffusion is a novel framework that effectively leverages diffusion models to capture the stochastic and multimodal nature of quadrotor dynamics. This work introduces a novel framework that leverages conditional diffusion models to learn quadrotor dynamics, formulated as a sequence generation task. DroneDiffusion achieves superior generalization to unseen, complex scenarios by capturing the temporal nature of uncertainties and mitigating error propagation.
Experiment 1: Generalization to Unseen Complex Trajectory
Task: Training data are collected by manoering the quadrotor on the primitive trajectories and then evaluated on the complex flight paths unseen during training.
Experiment 2 (A): Adaption against Payload and Velocities
Task: Training data are collected by manoering the quadrotor with the payloads {0.2, 0.6} kg and velocities {0.2, 0.4} m/s.
Evaluation: (i) Payload {0.4} kg and velocity {0.35} m/s are in distribution.
(ii) Payload {0.8} kg and velocity {0.50} m/s are out of distribution.
Experiment 2 (B): Adaption against Wind
Task: Stabilizing the quadrotor at (0, 0, 0.8) m under external wind disturbance.
Experiment 3: Runtime performance with multi-step predictions
Task: Stabilizing the quadrotor at (0, 0, 0.8) m for 60 sec for different horizon lengths. Sampling from the diffusion models are then used sequentially in the control law.
BibTeX
@article{das2024dronediffusion,
title={DroneDiffusion: Robust Quadrotor Dynamics Learning with Diffusion Models},
author={Das, Avirup and Yadav, Rishabh Dev and Sun, Sihao and Sun, Mingfei and Kaski, Samuel and Pan, Wei},
journal={arXiv preprint arXiv:2409.11292},
year={2024}
}