Source-linked AI summary
UWB Meets Crazyflow: Simulating Degraded Feedback at Scale for Aerial Robotics
Martin Schuck, Marcel P. Rath, Yufei Hua, Abhishek Goudar, SiQi Zhou, Angela P. Schoellig
TL;DR
Quadrotor simulation still faces a scaling bottleneck for realistic imperfect feedback, while differentiable frameworks may compromise model fidelity or demonstrated sim-to-real capability. Crazyflow combines JAX-based differentiable simulation with UWB/IMU sensing and EKF estimation. GPU throughput remains high with the EKF, while poorer sensor quality degrades tracking.
Problem
Evaluating controllers with realistic noisy state feedback across thousands of parallel rollouts remains difficult, and differentiable frameworks may compromise fidelity or sim-to-real capability.
Source: §I · p. 1Method
Crazyflow is a JAX-based differentiable simulator integrating physics, sensor pipelines, and controllers, demonstrated with UWB/IMU sensing and an EKF.
Source: §I · p. 1, §II · p. 1Results
GPU throughput falls roughly 20% with the 19-state EKF yet remains near ∼8 × 10^6 steps per second, while poorer UWB/IMU quality degrades EKF position estimates and tracking.
Source: §III · p. 2Takeaways & Limitations
The integrated pipeline remains differentiable, and the authors report that RL agents can be trained in seconds and deployed on physical hardware without complex domain randomization.
Source: §III · p. 2, §IV · p. 2
Abstract
from arXiv · showhide
In this work, we introduce Crazyflow, an accurate, differentiable simulator built on JAX. By leveraging jit compilation via XLA, Crazyflow unifies physics and control into a single differentiable computation graph, enabling massive parallelization on accelerated hardware without sacrificing modeling accuracy. This architecture achieves order-of-magnitude speedups over existing baselines, capable of training deployable reinforcement learning agents in seconds. To highlight its highly modular design, we demonstrate how easily Crazyflow can be extended by integrating a complete, high-fidelity Ultra-Wideband (UWB) and Inertial Measurement Unit (IMU) simulation pipeline coupled with a full-state Extended Kalman Filter (EKF). This capability allows for massive parallel controller evaluation under realistic, degraded state feedback with minimal impact on GPU throughput. By combining speed, accuracy, and extensibility, Crazyflow serves as a foundational tool for the next generation of aerial robotics research.
I. INTRODUCTION
Quadrotors need efficient simulation for rapid iteration and optimization, but evaluating controllers with realistic imperfect feedback at scale remains difficult. Crazyflow addresses this need with a differentiable simulator integrating physics, sensing, estimation, and control.
I. INTRODUCTION
- Quadrotor research needs high-performance simulation for rapid algorithmic iteration and task optimization across individual-robot and multi-agent settings.Applications range from infrastructure inspection and safety-critical tasks, to racing and shows,.Source: §I · p. 1
- Evaluating controllers under noisy UWB or vision feedback across thousands of parallel rollouts remains a scaling bottleneck for widely used simulators.RotorS, CrazyS, Crazyswarm, and CrazySim are generally not designed for massively parallel GPU execution.Source: §I · p. 1
- Scalable tensorized or distributed simulators typically lack analytical dynamics gradients, while differentiable alternatives may compromise fidelity or demonstrated sim-to-real capability.The latter limitation is reported outside simplified settings.Source: §I · p. 1
- Crazyflow integrates physics, modular sensing, and controllers in a differentiable JAX computation graph, demonstrated with UWB/IMU sensing and an EKF for imperfect-feedback evaluation at scale.The paper reports retaining high speed, direct sim-to-real transfer, and full differentiability.Source: §I · p. 1
II. THE CRAZYFLOW DESIGN
Crazyflow uses JAX to combine just-in-time compilation, parallel acceleration, and differentiability, dynamically composing the functions required by each configuration. Its focus on unconstrained free flight avoids a general-purpose solver step.
II. THE CRAZYFLOW DESIGN
- JAX provides Crazyflow with just-in-time compilation, massively parallel hardware acceleration, and full differentiability.The simulator dynamically composes only the functions required for the current configuration.Source: §II · p. 1
- Crazyflow dynamically composes its simulation pipeline using only functions required by the current configuration.The architecture is built around JAX and supports massively parallel hardware acceleration.Source: §II · p. 1
- By focusing on unconstrained free flight, Crazyflow omits the solver step used in general-purpose simulators, simplifying gradient tracing and enhancing performance.The source characterizes that solver step as computationally expensive.Source: §II · p. 1
A. Core Dynamics and Sim-to-Real
Crazyflow is modeled on the Crazyflie family and offers detailed first-principles and abstracted attitude dynamics models. Its system-identification pipeline supports adapting the simulator to other drone platforms.
A. Core Dynamics and Sim-to-Real
- Crazyflow provides a detailed first-principles dynamics model capturing nonlinear motor dynamics and an abstracted attitude model for standard mid-level control interfaces.The models are natively based on the Crazyflie quadrotor family.Source: A. Core Dynamics and Sim-to-Real · p. 1
- A streamlined system-identification pipeline lets users extract and match physical parameters for new drone platforms.The paper presents this as a way to adapt the simulation framework beyond its Crazyflie-native models.Source: A. Core Dynamics and Sim-to-Real · pp. 1–2
- The authors state that adapting platform parameters supports reliable sim-to-real transfer regardless of the hardware.Source: A. Core Dynamics and Sim-to-Real · p. 2
B. Extensibility: Sensing and Estimation
Crazyflow integrates sensing and state estimation into simulator-step execution before the controller. Its example pipeline combines noisy IMU and UWB measurements in a full-state EKF, allowing controller inputs to use either estimated or ground-truth state.
B. Extensibility: Sensing and Estimation
- The stateless, pure-function architecture permits customization, with sensing and estimation integrated into simulator-step execution ahead of the controller.Source: B. Extensibility: Sensing and Estimation · p. 2
- The IMU models noisy body-frame specific force and angular velocity, while UWB ranges use one drone tag and eight fixed anchors under assumed line-of-sight conditions.UWB range noise has fixed standard deviation σ and per-anchor mean µi sampled at initialization and held constant per episode.Source: B. Extensibility: Sensing and Estimation · p. 2
- A full-state EKF tracks 19 states, predicts from IMU-driven dynamics with Jacobian-based covariance propagation, and corrects with UWB ranges.Controllers can switch between perfect ground-truth state and estimated feedback.Source: B. Extensibility: Sensing and Estimation · p. 2
III. EVALUATION RESULTS
The evaluation tests throughput with integrated state estimation and closed-loop tracking under varied sensor quality. GPU throughput falls modestly, while lower-quality sensing worsens estimates and tracking.
III. EVALUATION RESULTS
- A roughly 20% throughput reduction across parallel-world counts leaves GPU peak performance at ∼8 × 10^6 steps per second versus slightly over 10^7 for baseline.The throughput comparison is shown for simulations with and without the full state-estimation pipeline.Source: Figure 2 · p. 2, §III · p. 2
- Lower-quality UWB+IMU sensing degrades the EKF position estimate and closed-loop figure-8 tracking.The geometric controller tracks a figure-8 reference with a 10 s cycle time under ground-truth, high-quality, and low-quality feedback conditions.Source: Figure 3 · p. 2, §III · p. 2
- Despite sensor noise, bias sampling, and EKF linearizations, the full pipeline remains differentiable, preserving BPTT-based RL training and direct hardware deployment.Source: §III · p. 2
IV. CONCLUSION
Crazyflow combines speed, differentiability, and physical realism, with integrated UWB/IMU sensing and EKF state estimation for scalable evaluation under imperfect feedback.
IV. CONCLUSION
- Crazyflow integrates a full UWB/IMU and EKF pipeline into simulation steps with minimal GPU overhead for robust-control evaluation under imperfect feedback at scale.Source: §IV · p. 2
- An integrated system-identification pipeline lets researchers adapt Crazyflow to any drone platform.Source: §IV · p. 2
- Crazyflow supports training deployable RL agents in seconds and reliable, direct sim-to-real transfer.Source: §IV · p. 2