Title: Crazyflow: A High-Precision, GPU-Enabled, Differentiable Drone Simulator Built on JAX

Abstract:

The generation of high-fidelity, large-scale synthetic data through simulation has emerged as a critical foundation for advancing robotic algorithms. Although current aerial robotics simulators have developed specialized capabilities—such as high fidelity, differentiability, and swarm support—there remains a lack of a unified platform capable of synthesizing data across all these domains simultaneously. To address this gap, we introduce Crazyflow, a simulator engineered to expand the boundaries of aerial robotics algorithm development. It accommodates a wide spectrum of methodologies, ranging from model-based to data-driven techniques, gradient-based to sampling-based strategies, and single-agent to multi-agent configurations.

In terms of performance, Crazyflow surpasses existing state-of-the-art drone simulators by more than an order of magnitude in speed for individual units, while also demonstrating the capacity to simulate thousands of swarms, each comprising 4,000 drones. Empirical results from real-world experiments indicate that Crazyflow facilitates analytical-gradient-based policy learning, achieving trajectory tracking accuracy within sub-centimeter ranges without the need for domain randomization. Additionally, it supports sampling-based obstacle avoidance at rates exceeding 500 million steps per second.

Challenging the conventional "train-then-deploy" model, the simulator’s exceptional speed allows for reinforcement learning to occur in mid-flight. We validated this capability by launching a physical drone and training a recovery policy from scratch in just 0.38 seconds, which successfully stabilized the aircraft. Crazyflow offers various levels of simulation abstraction and maintains direct compatibility with all open-source Crazyflie models. It also streamlines adaptation for custom drone platforms and applications through a lightweight system identification pipeline. By simultaneously enhancing accuracy, speed, and differentiability, Crazyflow provides an open-source solution for synthetic data generation, featuring emerging large-scale parallelization capabilities that support online, in-execution learning and optimization, thereby paving the way for novel algorithmic developments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC