Automatic Generation of Aerobatic Flight in Complex Environments via Diffusion Models

TL;DR

The paper tackles the challenge of long-horizon aerobatic flight in cluttered environments by learning a library of aerobatic primitives and composing them with diffusion models. It introduces AeroDM, a conditional diffusion transformer that generates primitives conditioned on target waypoints and action semantics, while incorporating historical priors and collision guidance to maintain continuity and safety. A hierarchical post-processing stage based on MINCO trajectory optimization then enforces dynamic feasibility and safety corridors, enabling real-world deployment. The approach is validated through extensive simulations and real-world flights, showing high success rates, controllability via conditioning, and robust obstacle handling, with practical implications for scalable, editable, long-horizon drone aerobatics. The work advances autonomous, visually striking flight in complex environments and sets the stage for scene-aware extensions that adapt maneuvers to environmental features.

Abstract

Performing striking aerobatic flight in complex environments demands manual designs of key maneuvers in advance, which is intricate and time-consuming as the horizon of the trajectory performed becomes long. This paper presents a novel framework that leverages diffusion models to automate and scale up aerobatic trajectory generation. Our key innovation is the decomposition of complex maneuvers into aerobatic primitives, which are short frame sequences that act as building blocks, featuring critical aerobatic behaviors for tractable trajectory synthesis. The model learns aerobatic primitives using historical trajectory observations as dynamic priors to ensure motion continuity, with additional conditional inputs (target waypoints and optional action constraints) integrated to enable user-editable trajectory generation. During model inference, classifier guidance is incorporated with batch sampling to achieve obstacle avoidance. Additionally, the generated outcomes are refined through post-processing with spatial-temporal trajectory optimization to ensure dynamical feasibility. Extensive simulations and real-world experiments have validated the key component designs of our method, demonstrating its feasibility for deploying on real drones to achieve long-horizon aerobatic flight.

Figures (8)

• Figure 1: Illustration of aerobatic primitive generation, the trajectory segments containing the aerobatic maneuver are segmented, and the discretized motion sequences are sampled from it. Redundant trajectory segments are added to simulate the transition between aerobatic primitives.
• Figure 2: The architecture of the diffusion process. (A) Schematic of the overall process. (B) Detailed structure of the Aerobatic Diffusion Model.
• Figure 3: Five different maneuver styles of aerobatic trajectories: (a) the Power Loop, (b) the Barrel Roll, (c) the Split-S, (d) the Immelmann Turn, (e) the Wall Ride.
• Figure 4: Up: results of the aerobatic generation conditioned on the "Power Loop" action (a) compared with action-agnostic model (b). Down: box plot of the errors between the terminal of aerobatic primitives and the given target, measured by distances.
• Figure 5: Comparison between models with and without access to previous primitives. The smoothness is measured with differences of positions $\delta p$ and Euler angles $\delta \theta$ between adjacent frames.