A JAX-based quadrotor simulator that runs batched, differentiable drone simulations on CPU and GPU for Crazyflie 2.x drones.
Crazyflow is a research simulator for quadrotor drones. It is built on JAX and runs many independent environments and multi-drone swarms in parallel on CPU or GPU, which makes it suited to large-scale training and experiments in robot learning and control.
The simulator offers two kinds of dynamics for the Crazyflie 2.x family: first-principles equations with parameters identified from real-world measurements, and simplified abstracted dynamics in three variants fitted from real flight data. Because the pipeline is differentiable, gradients can flow through the full dynamics and control chain with jax.grad. Step and reset functions are plain JAX functions that users can extend or swap out.
MuJoCo integration via MJX provides onscreen and offscreen rendering, raycasting, and contact detection.
Install the CPU version from PyPI
pip install crazyflowInstall the GPU version (Linux x86-64, CUDA 12)
pip install "crazyflow[gpu]"Read the documentation for examples and the user guide
Summary generated by AI from the project's README. GitHub stats refresh daily.
A JAX-based rigid-body physics engine that runs massively parallel, differentiable simulation on accelerators for robotics and RL research.
PyBullet-based Gymnasium environments for training single and multi-agent reinforcement learning controllers for quadcopters.
A Python framework that JIT-compiles regular Python functions into kernels running on CPU or NVIDIA GPU, for simulation, robotics, and ML.
A framework on NVIDIA Isaac Sim for simulating multirotor drones, with PX4 and ArduPilot autopilot integration and a Python control interface.
A Unreal Engine-based simulator for drones, cars and other vehicles, extended by Cosys-Lab with new sensor models and APIs.
An imitation learning benchmark for whole-body control, with MuJoCo and MJX environments for humanoids, quadrupeds and musculoskeletal human models.