Title: DiffAero: A GPU-Accelerated Differentiable Simulation Framework for Efficient Quadrotor Policy Learning

Abstract:

This publication presents DiffAero, a streamlined, fully differentiable simulation environment accelerated by GPUs, specifically engineered to streamline the learning of quadrotor control policies. DiffAero is capable of supporting both agent-level and environment-level parallelism. It features a unified, GPU-native training interface that incorporates various dynamics models, customizable sensor configurations (including LiDAR, depth cameras, and IMUs), and a range of flight tasks. By offloading both physics calculations and rendering entirely to the GPU, the framework removes the latency associated with CPU-GPU data transfers, resulting in simulation throughput that is orders of magnitude higher than previous methods. Unlike current simulators, DiffAero functions not only as a high-performance simulation tool but also as a research platform for investigating hybrid and differentiable learning algorithms. Real-world flight tests and comprehensive benchmarks confirm that combining DiffAero with hybrid learning algorithms enables the acquisition of robust flight policies within hours on standard consumer hardware. The source code can be accessed at https://github.com/flyingbitac/diffaero.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC