Source-linked AI summary
Agilicious: Open-Source and Open-Hardware Agile Quadrotor for Vision-Based Flight
Philipp Foehn, Elia Kaufmann, Angel Romero, Robert Penicka, Sihao Sun, Leonard Bauersfeld, Thomas Laengle, Giovanni Cioffi, Yunlong Song, Antonio Loquercio, Davide Scaramuzza
TL;DR
Autonomous agile quadrotor research needs a platform that combines high maneuverability with sufficient onboard sensing, computation, and flexible software. Agilicious addresses this gap through co-designed open hardware and modular software, and its experiments demonstrate use across agile-flight research tasks and deployment settings.
Problem
Autonomous agile flight requires simultaneous onboard computation, neural-network acceleration, agile actuation, and flexible software, while existing platforms satisfy only subsets of these requirements.
Method
Agilicious co-designs an open-source, open-hardware quadrotor with high-performance actuation, onboard GPU computing, and modular software that transfers algorithms from simulation to real-world deployment.
Results
The platform supports experiments in agile control, trajectory planning, learning-based research, latency evaluation, vision-based state estimation, and research demonstrators across simulation and real-world settings.
Takeaways & Limitations
Agilicious provides a standardized platform for combining and replacing hardware and software components while evaluating autonomous agile-flight algorithms.
Takeaways & Limitations
The platform must simultaneously carry sufficient compute hardware for autonomous operation and retain the thrust-to-weight and torque-to-inertia ratios needed for agile flight.
Abstract
from arXiv · showhide
Autonomous, agile quadrotor flight raises fundamental challenges for robotics research in terms of perception, planning, learning, and control. A versatile and standardized platform is needed to accelerate research and let practitioners focus on the core problems. To this end, we present Agilicious, a co-designed hardware and software framework tailored to autonomous, agile quadrotor flight. It is completely open-source and open-hardware and supports both model-based and neural-network--based controllers. Also, it provides high thrust-to-weight and torque-to-inertia ratios for agility, onboard vision sensors, GPU-accelerated compute hardware for real-time perception and neural-network inference, a real-time flight controller, and a versatile software stack. In contrast to existing frameworks, Agilicious offers a unique combination of flexible software stack and high-performance hardware. We compare Agilicious with prior works and demonstrate it on different agile tasks, using both model-based and neural-network--based controllers. Our demonstrators include trajectory tracking at up to 5g and 70 km/h in a motion-capture system, and vision-based acrobatic flight and obstacle avoidance in both structured and unstructured environments using solely onboard perception. Finally, we demonstrate its use for hardware-in-the-loop simulation in virtual-reality environments. Thanks to its versatility, we believe that Agilicious supports the next generation of scientific and industrial quadrotor research.
1. INTRODUCTION
Agilicious addresses the competing demands of autonomous agile flight by co-designing high-performance hardware with open, modular software. Its standardized platform combines substantial onboard computation, agile flight capability, and flexible deployment across simulation and real-world experiments.
- Motivation: Autonomous agile flight supports time-critical missions but requires platforms that combine sensing, computation, and high maneuverability.Existing platforms often emphasize either heavy autonomous sensor and compute setups or agility without sufficient resources for autonomy.
- Design requirements: The platform co-design meets autonomous-flight requirements through onboard estimation, planning, control, neural-network acceleration, and agile thrust-to-weight and torque-to-inertia ratios.Increasing platform size and weight can improve thrust but typically reduces torque-to-inertia ratio because inertia grows quadratically with size and linearly with weight.
- Limitations of prior platforms: Existing platforms trade off autonomy and agility, with limited actuation, insufficient compute, or proprietary low-level controllers constraining research.These limitations reduce control over flight characteristics and hinder interpretation, modeling, and exploitation of aerodynamic and high-frequency effects.
- Platform contribution: Agilicious is an open-source and open-hardware platform designed for more than 5g acceleration while providing substantial onboard computation and versatile software.Its lightweight 750 g hardware reaches a maximal speed of 131 km/h and includes GPU-accelerated onboard computing for neural-network inference and optimization-based control.
- Software architecture: Agilicious uses a modular software architecture to transfer algorithms from simulation to real-world deployment and onboard sensing in unknown environments.Components can be rapidly substituted and reused in standalone tests, experiments, benchmarks, and other applications.
- Validation and versatility: The standardized hardware and software platform has been tested across diverse environments with different estimation, control, and planning strategies, facilitating new algorithm evaluation.The framework supports simulation, real-world experiments, hardware-in-the-loop, and rendering engines through adaptable interfaces.
2. RESULTS
The experiments evaluate Agilicious for agile control, trajectory planning, and learning-based research across simulation and real-world settings. They examine tracking, pipeline latency, vision-based state estimation, and demonstrators built on the platform.
- Experimental scope: Agilicious experiments span agile control, trajectory planning, learning-based research, latency evaluation, and vision-based state estimation.The evaluation combines simulation, real-world experiments, and two demonstrators from recent research projects.
A. Trajectory Tracking Performance
Agilicious evaluates aggressive trajectory tracking across onboard, offboard, and simulated configurations. The custom agiNuttx controller achieves the strongest real-world tracking, while simulation closely reproduces flight behavior.
- Control configurations: Onboard agiNuttx combines MPC with INDI to compute refined single-rotor thrust commands and execute them through closed-loop motor-speed tracking.This configuration is designed to reduce sensitivity to model inaccuracies.
- Control configurations: Offboard execution streams bodyrate and collective-thrust commands wirelessly, enabling computationally demanding algorithms but adding latency that can degrade control performance.The offboard setup runs the Agilicious stack on a desktop or laptop computer.
- Tracking results: 0.322 m average positional RMSE is achieved by onboard computation with agiNuttx, outperforming BetaFlight at less than 0.385 m and offboard control at 0.474 m.The offboard configuration incurs higher latency, while simulation reaches 0.320 m RMSE.
- Simulation: Simulation closely matches real-world tracking characteristics, supporting transition from simulation prototyping to real-world verification without source-code changes.The simulation uses accurate quadrotor, motor-dynamics, and blade-element momentum aerodynamic models.
B. Control Latency
The platform analyzes command-to-actuation latency across agiNuttx, BetaFlight, and offboard control configurations. These delays are especially relevant when tracking agile trajectories under model mismatch, disturbances, and actuator limits.
- Latency evaluation: Command-to-actuation latency is analyzed for agiNuttx, BetaFlight, and BetaFlight with offboard control.The experiment measures how each low-level configuration responds to thrust commands.
- Measurement method: 12 N collective-thrust steps are measured with a load cell and fitted through a first-order motor-dynamics model.Force and moment measurements are time-synchronized with the commands.
C. Visual-Inertial State Estimation
Agilicious compares an integrated stereo VIO system with a monocular-camera SVO pipeline using the flight-controller IMU and onboard compute. Both perform well, while the integrated RealSense achieves lower error.
- Evaluation: Two VIO solutions are evaluated against ground truth on circle and Lemniscate trajectories flown with Agilicious.The circle has a 4 m radius at 5 m/s, while the Lemniscate has 5 m amplitude and speeds up to 7 m/s.
- Results: 0.151 m and 0.114 m ATE RMSE are achieved by Intel RealSense on the circle and Lemniscate, compared with 0.217 m and 0.131 m for monocular SVO.RealSense uses stereo vision and IMU; SVO uses one camera, the flight-controller IMU, and Jetson TX2 compute.
- Scope boundary: The Intel RealSense T265 is being discontinued, requiring consideration of alternative cameras, integrated sensors, or open-source VIO frameworks.The paper lists several possible alternatives.
D. Demonstrators
Agilicious is presented as a flexible research platform through hardware-in-the-loop simulation and autonomous flight using only onboard sensing. These demonstrations cover diverse research settings and applications.
- Demonstrators: The demonstrators cover hardware-in-the-loop simulation and autonomous flight in the wild using only onboard sensing.They are intended to illustrate the broad applicability of the Agilicious software and hardware stack.
D.1. Hardware in the Loop Simulation
Agilicious supports hardware-in-the-loop simulation to develop vision-based navigation in complex environments without risking crashes into real obstacles. Its high-fidelity simulator provides configurable sensors with delays of 13–22 ms, enabling rapid system development.
- Hardware-in-the-loop simulation combines real-world quadrotor dynamics and proprioceptive sensors with virtual photorealistic environments.This avoids idealized virtual devices while allowing arbitrarily sparse or dense environments without crashing into real obstacles.
- The Unity-based simulator provides RGB, depth, optical-flow, and segmentation sensors with configurable noise, blur, distortion, wind, and rain.
- 13–22 ms delays support high-fidelity rendering from 640×480 VGA to 1920×1080 full HD on an NVIDIA RTX 2080 GPU.
- Integrating the agile platform with visual simulation enables rapid development of vision-based navigation in complex, unstructured environments.
D.2. Vision-based Agile Flight with Onboard Sensing and Compute
Agilicious targets the latency and robustness challenges of onboard vision during fast flight and supports quantitative comparisons of traditional and learning-based navigation methods. These comparisons cover acrobatic maneuvers and high-speed obstacle avoidance in structured and unstructured environments.
- Onboard vision must remain effective, low-latency, and robust because motion blur, large pixel displacements, and illumination changes can cause failures at high speeds.
- Agilicious compares traditional and learning-based methods on agile navigation tasks in structured and unstructured environments.
- The framework supports quantitative analyses of autonomous acrobatic maneuvers and high-speed obstacle avoidance in previously unknown environments.
3. DISCUSSION
The discussion presents Agilicious as an open, versatile platform combining advanced computing with agile flight hardware. It identifies future directions in miniaturization, dynamic morphology, and faster, more robust vision-based autonomy.
- Agilicious combines advanced computing with an open-source and open-hardware quadrotor platform for autonomous-robot research.
- Future flying robots may become smaller, lighter, cheaper, and less power-consuming while improving battery life, crash resilience, and agility ratios.
- The framework was mainly demonstrated with fixed-shape quadrotors, which are easier to model and control and less susceptible to hardware failures.
- Dynamic-morphology platforms offer greater environmental adaptability and potentially higher power efficiency, including transformation from quadrotor to fixed-wing aircraft.
- Vision-based agile flight remains below professional human performance, with challenges in aerodynamics, low-latency perception, state estimation, and high-speed failure recovery.
- Agilicious provides a modular, open testbed that supports varied autonomous tasks and integration of novel sensors or compute hardware.
4. MATERIALS AND METHODS
Agilicious was designed through hardware–software co-design that balances competing objectives, including onboard computing capacity and platform agility.
- The platform co-design explains the flight hardware, compute hardware, and software choices underlying Agilicious.
A. Compute Hardware
Agilicious combines real-time low-level control with GPU-capable, general-purpose onboard computing and lightweight hardware for autonomous agile flight. Its sensing and compute choices support fast prototyping, neural-network inference, and visual-inertial navigation.
- Low-Level Flight Controller: Real-time low-level control stabilizes high-bandwidth dynamics that general-purpose operating systems cannot reliably handle.The controller has limited compute capability but provides reliable real-time performance for motor-speed and bodyrate control.
- High-Level Compute Board: Linux was selected to support broad software compatibility for general flight applications, rapid prototyping, and experimentation.CUDA support in Nvidia products further favors the selected compute modules for hardware-accelerated neural-network inference.
- High-Level Compute Board: The Jetson TX2 is preferred among viable CUDA-enabled modules because of adapter-board availability, interface diversity, and smaller footprint.The recommended ConnectTech Quasar breakout board provides USB, Ethernet, serial, and other interfaces for sensors and cameras.
- Low-Level Flight Controller: Agilicious provides its own autonomous-flight controller, agiNuttx, because BetaFlight targets human-piloted drones and exposes sensor readings only at 10 Hz.agiNuttx reuses compatible commercial flight-controller hardware while providing real-time low-latency interfacing and control.
- Airframe and Propulsion: The lightweight design uses off-the-shelf components while accommodating the Jetson TX2 compute unit and propulsion hardware.The selected motor-propeller combination provides continuous static thrust of 4× 9.5 N on the quadrotor.
- Sensing: Visual-inertial odometry is suited to arbitrary uninstrumented environments because cameras and IMUs offer complementary sensing with low price, availability, and useful depth information.The platform supplies high-rate filtered inertial measurements and can combine them with off-the-shelf cameras and VIO software.
D. The Agilicious Flight Stack Software
The Agilicious flight stack is an open-source, modular system designed for low-latency agile flight, flexible deployment, and transfer between simulation and real platforms. It combines reusable pipeline modules, hardware interfaces, simulation, and safety backup configurations.
- Flight Stack: The open-source flight stack targets high versatility, modularity, low latency for agile flight, and transferability across simulation and multiple real-world platforms.Its architecture is split into a core library and a ROS wrapper.
- Core Library: agilib provides minimally dependent modules that run from lightweight devices to heterogeneous server architectures.Avoiding dependencies reduces compatibility issues across computing platforms.
- ROS Wrapper: agiros adds ROS compatibility, network communication, logging, a GUI, and integration with other software while preserving deployment options without ROS.Separating agiros from agilib avoids unnecessary communication overhead where ROS is unavailable or not needed.
- Pipeline: Each pipeline control cycle sequentially runs an estimator, sampler, controllers, and bridge to track a reference, while modules may process asynchronous tasks in parallel.The pilot orchestrates these modules through a unified pipeline structure.
- Estimator: The estimator produces time-stamped state estimates, with feed-through and extended Kalman-filter implementations that can be replaced or extended for additional sensors and localization pipelines.The state includes position, orientation, velocity, bodyrate, accelerations, higher derivatives, biases, and rotor thrusts.
- Sampler and Controller: The sampler supplies trajectory-progress setpoints using either elapsed time or current position before the controller generates closed-loop commands.Available controllers include a full nonlinear-model MPC and a cascaded geometric controller.
- Bridge: The bridge sends rotor-thrust or bodyrate commands through interfaces such as ROS, SBUS, and serial.This abstracts communication with low-level controllers and other hardware or software sinks.
- Simulator: The simulator models motor dynamics, aerodynamics, and optionally low-level-controller behavior to support high-fidelity prototyping and testing.It can also support simulation-based continuous-integration testing without external framework dependencies.