PaperlayerView exact revision on arXiv

Source-linked AI summary

The MRS UAV System: Pushing the Frontiers of Reproducible Research, Real-world Deployment, and Education with Autonomous Unmanned Aerial Vehicles

Tomas Baca, Matej Petrlik, Matous Vrba, Vojtech Spurny, Robert Penicka, Daniel Hert, Martin Saska

arXiv:2008.08050v6cs.ROeess.SY

TL;DR

Research UAVs need reproducible, realistic full-stack systems that remain reliable across diverse sensors, environments, and flight conditions. This paper presents an open-source multirotor control and estimation platform using multi-frame localization, heading-based orientation, and complementary feedback controllers. The system supports real-world deployment, including fast maneuvers and flight with noisy or unreliable sensors, and has been used in demanding robotics competitions.

  • Problem

    Only a handful of UAV platforms support research out of the box, while replicable applied robotics requires realistic simulation and real-world experimental verification.

  • Method

    The platform combines multi-frame multi-sensor localization, rotation-matrix and heading-based orientation, two feedback controllers, disturbance estimation, modular software, and realistic simulation.

  • Results

    The open-source system supports reliable deployment outside laboratories, fast and agile maneuvers, safe flight with noisy or unreliable sensors, and real-world robotics-competition use.

  • Takeaways & Limitations

    The actively maintained platform provides a reusable basis for research, development, testing, and multi-robot aerial applications.

  • Takeaways & Limitations

    The cascade control design assumes an exponentially stable inner attitude loop with bandwidth greater than the outer-loop dynamics, and one orientation structure does not preserve heading generally.

Abstract

from arXiv · show

We present a multirotor Unmanned Aerial Vehicle control (UAV) and estimation system for supporting replicable research through realistic simulations and real-world experiments. We propose a unique multi-frame localization paradigm for estimating the states of a UAV in various frames of reference using multiple sensors simultaneously. The system enables complex missions in GNSS and GNSS-denied environments, including outdoor-indoor transitions and the execution of redundant estimators for backing up unreliable localization sources. Two feedback control designs are presented: one for precise and aggressive maneuvers, and the other for stable and smooth flight with a noisy state estimate. The proposed control and estimation pipeline are constructed without using the Euler/Tait-Bryan angle representation of orientation in 3D. Instead, we rely on rotation matrices and a novel heading-based convention to represent the one free rotational degree-of-freedom in 3D of a standard multirotor helicopter. We provide an actively maintained and well-documented open-source implementation, including realistic simulation of UAV, sensors, and localization systems. The proposed system is the product of years of applied research on multi-robot systems, aerial swarms, aerial manipulation, motion planning, and remote sensing. All our results have been supported by real-world system deployment that shaped the system into the form presented here. In addition, the system was utilized during the participation of our team from the CTU in Prague in the prestigious MBZIRC 2017 and 2020 robotics competitions, and also in the DARPA SubT challenge. Each time, our team was able to secure top places among the best competitors from all over the world. On each occasion, the challenges has motivated the team to improve the system and to gain a great amount of high-quality experience within tight deadlines.

1 INTRODUCTION

Multirotor UAV research still requires substantial full-stack development, realistic simulation, and real-world verification. The proposed platform addresses these needs with an actively maintained, modular open-source system spanning estimation, control, simulation, and deployment.

  • Only a handful of UAV platforms support research out of the box, leaving researchers to develop full-stack control and guidance.
  • Replicating and validating applied robotics research requires experimental platforms together with realistic simulation environments and supporting materials.
  • The platform combines indoor and outdoor operation, onboard multi-sensor fusion, realistic simulation, and two controllers for aggressive or stable flight.
  • Its contributions include multi-frame localization, heading-oriented control without Euler/Tait-Bryan angles, estimator-independent disturbance handling, and MPC-based force feedback.
  • The modular system supports varied sensors and localization systems, coordinate-frame-separated references, fast waypoint trajectories, and mid-flight hot-swapping of methods.
  • Real-world testing covered outdoor fields, forests, factories, mines, caves, tunnels, manipulation, aggressive flight, moving-platform landing, and multi-UAV swarming.

2 Multirotor aerial vehicle dynamics model

The UAV model represents position and orientation using world and body frames, a rotation matrix, and angular velocity. It defines heading from the projected body x-axis rather than relying on yaw or Tait-Bryan angles.

  • State representation: The UAV state comprises world-frame position, velocity, acceleration, body-to-world rotation matrix, and body-frame angular velocity.The rotation matrix R belongs to SO(3), satisfies det R = 1 and Rᵀ = R⁻¹, and angular velocity obeys Ωv = ω × v.
  • Dynamics: The translational dynamics include downward gravity and collective propeller thrust directed along the body-frame b̂3 axis.The model separately considers the azimuth of b̂1 as the UAV heading for non-aerobatic flight.
  • Heading convention: Heading is defined as the azimuth of the body-frame b̂1 axis in the world frame under a nonvertical projection condition.The definition requires |ê3ᵀb̂1| < 1.
  • Heading convention: Heading provides a more intuitive controllable degree of freedom than yaw, while yaw is reliable as an approximation only near-horizontal orientation.The paper advises against Tait-Bryan angles because conventions vary and yaw does not directly identify a particular body-axis orientation independently of roll and pitch.

3 State estimation

State estimation uses separated translational and rotational components, a simple high-level model, and multiple estimator hypotheses for diverse sensors and reference frames. An arbiter selects the current estimate while preserving smooth control during switches.

  • Estimator architecture: The estimator separates high-level translational-state estimation from attitude and angular-rate estimation across onboard computing and the embedded flight controller.The translational submodel estimates position, velocity, and acceleration, while attitude estimation uses the flight controller and IMU.
  • Estimator model: The high-level state vector contains position, its first two derivatives, heading, and heading rate.The state is x = [x, ẋ, ẍ, y, ẏ, ÿ, z, ż, z̈, η, η̇]ᵀ.
  • Estimator model: A decoupled discrete linear time-invariant model uses four independently estimated subsystems and only three free transfer parameters.Assuming identical horizontal axes simplifies tuning and reuse across UAV platforms; decoupling also speeds computation.
  • Sensor integration: Sensors providing position, heading, horizontal position, height, or velocity can be incorporated through corresponding measurement configurations.Examples include 3D SLAM, GNSS with a rangefinder, 2D laser SLAM, and optic-flow systems.
  • Estimator design: The translational estimator uses a recursive linear Kalman filter whose stable, fast tracking depends on controller compensation for force and input biases.The authors favor this pragmatic separation because it permits replacement by other state sources and keeps the platform universal.
  • Multiple hypotheses: A bank of Kalman filters estimates UAV states from different sensor subsets instead of combining every measurement into one hypothesis.The filters support transitions between GNSS and indoor SLAM, parameter variants, and backup estimators for emergency landing.
  • Multiple hypotheses: The arbiter can switch estimates when requested, when the current filter becomes unreliable, or when another hypothesis has the smallest covariance trace.Internal tracker and controller states are recalculated so the switch remains smooth to an outside observer.
  • Multiple hypotheses: The system maintains hypotheses in different frames and selects the best current estimate while allowing feedback control and reference generation in other frames.This supports heterogeneous sources such as visual SLAM and GPS without requiring a common measurement frame.

4 Feedforward tracking and reference generation

Feedforward tracking converts position-heading references or trajectories into smooth, feasible full-state references for feedback control. Linear MPC generates constrained reference transients, with specialized handling for take-off, landing, and swarming.

  • Reference tracking: The Reference tracker accepts a 3D position and heading target or a time-indexed trajectory from the Mission & navigation block.Its output feeds a Feedback controller within the control pipeline.
  • Reference tracking: The control reference includes position, velocity, acceleration, jerk, heading, and heading rate for the differentially-flat translational dynamics.The represented components are x, ẋ, ẍ, x⃛, y, ẏ, ÿ, y⃛, z, ż, z̈, z⃛, η, and η̇.
  • MPC tracking: Linear MPC controls an ideal virtual UAV model in real time and samples its states as references for the feedback controller.The tracker produces optimal state transients while satisfying state constraints, at 100 Hz when generating the full-state reference χ.
  • Specialized tracking: Take-off and landing trajectories are generated separately to improve safety during phases where foliage or platform motion can create rapid control errors.The take-off solution uses an admittance-based approach to address entanglement-related disturbances.
  • Specialized tracking: The swarming tracker can bypass selected desired states and specify only the horizontal or vertical states required by the swarming mechanism.This accommodates approaches that command desired velocity or acceleration instead of classical horizontal-state tracking.

5 Feedback Control

The pipeline uses cascade-based feedback control with two controller variants: an agile SE(3)-based design and an MPC–SO(3) design for noisier estimates. It also adds heading-compliant orientation construction, disturbance compensation, thrust modeling, and failsafe-related mechanisms.

  • Controller variants: The agile controller extends SE(3) geometric feedback for fast, precise maneuvers when state estimates and references are smooth and model-consistent.
  • Controller variants: The second controller combines linear MPC with nonlinear SO(3) force tracking to provide more stable flight with noisy state estimates.
  • Pipeline architecture: The cascade architecture separates the attitude inner loop from the outer control loop under a time-scale-separation assumption.The inner loop is assumed exponentially stable and faster than the outer-loop dynamics.
  • Heading-compliant orientation: The proposed desired-orientation construction preserves the desired heading while aligning the desired force with the vehicle’s ˆb3d axis.The original construction does not generally preserve heading during maneuvers.
  • Heading control: Heading-rate feedforward and parasitic heading-rate compensation address heading changes induced by dynamic force-tracking maneuvers.The compensation is intended to act before feedback corrections appear during aggressive maneuvers.
  • Disturbance and thrust compensation: The pipeline estimates world- and body-frame disturbances, compensates apparent mass changes, and converts desired thrust force into collective motor speed.The thrust model uses an empirically fitted quadratic relationship between thrust and motor angular rate.

5.7 Model Predictive Control Force Feedback

The MPC force-feedback controller generates desired acceleration from a linear model while leaving heading control to SO(3) feedback.

  • Force feedback: The MPC controller generates a desired acceleration cd ∈R3, which is then used to calculate the desired force.

5.7.1 MPC Model

The MPC model predicts translational dynamics over a constrained horizon and solves a quadratic program at each control iteration. The resulting controller is designed to suppress estimator noise and enforce state and input constraints.

  • Model formulation: The MPC uses an LTI translational model and excludes heading from the prediction model because heading remains controlled through SO(3) feedback.
  • Model formulation: The model uses a sampling interval of ∆t := 0.05 s and applies the system input directly to the acceleration state without delay.
  • Optimization problem: The controller minimizes squared state-tracking errors over a prediction horizon while enforcing model, state, and input-slew constraints.The input slew-rate limits constrain system jerk.
  • References and initialization: The reference state holds desired positions with zero derivatives, while the initial condition can substitute feedforward reference derivatives when estimated states violate dynamic constraints.
  • Online control: The MPC solves its optimization problem at 100 Hz and extracts the acceleration reference from the first control input.
  • Operational choice: The controller is designed to be more stable with sensor noise and to satisfy state constraints intrinsically, so it is the pipeline’s default choice.The SE(3) controller is preferred when high accuracy is needed and localization is sufficiently smooth.

6 Trajectory generation

The trajectory generator extends waypoint-based polynomial optimization by selectively rescaling violating segments and estimating lower-bound segment times to avoid unnecessary flight-time increases.

  • Trajectory generation: The method generates real-time reference trajectories from waypoints while satisfying state constraints.
  • Constraint handling: It rescales only trajectory segments that directly violate constraints instead of rescaling the entire trajectory.
  • Time allocation: Lower-bound initial segment-time estimates prolong segments when necessary without shortening them when possible, improving overall flight time.

7 Implementation

The implementation provides a modular, ROS-based UAV control pipeline supporting estimator, tracker, controller, gain, and constraint switching during flight. Its open-source architecture is designed to transfer methods between realistic simulation and real hardware.

  • Modularity: The system supports hot-swapping feedback controllers, trajectory trackers, state estimators, gains, and dynamic constraints during flight.This modularity supports new methods, backup systems, and recovery from unwanted behavior.
  • State estimation: A bank of estimators provides simultaneous multi-frame localization from different subsets of sensor inputs.The design avoids forcing all available sensory inputs into a single localization hypothesis.
  • Feedback control: The controller bank combines an agile SE(3) controller for aggressive maneuvers with an MPC controller designed for estimation noise and disturbances.The SE(3) controller supports accelerations approaching 10 m s−2, while MPC handles noisy estimates and infeasible references more robustly.
  • Configuration: Global management of dynamic-constraint groups and estimator-specific gains allows speed, acceleration, jerk, snap, and rotational limits to change during flight.Gain groups are assigned to estimator types because different sensor-fusion systems have different noise characteristics.
  • System integration: The ROS implementation includes tracker and controller interfaces, synchronization, safety functions, take-off and landing state machines, and open-source documentation.The architecture also supports application-specific mission control while preserving the core pipeline.
  • Simulation-to-reality: Real and simulated flights show minimal differences, supporting a smooth transition between simulation and reality.The authors report that this substantially accelerates deployment of new robotic methods and algorithms.

8 Experimental evaluation

Experiments on a Tarot 650 platform compared the proposed simulator with real-world flights across step and circular references. The system maintained precise tracking near physical limits and provided robust control under localization disturbances.

  • Experimental setup: Real and simulated experiments on the Tarot 650 platform showed nearly identical system dynamics.The evaluation was conducted both in the proposed simulator and in real-world flights.
  • Step responses: Step references demonstrated precise and agile three-axis control near physical limits under velocity, acceleration, jerk, and snap constraints.The reported limits include 9 m s−1 velocity, 12 m s−2 acceleration, and 50 m s−3 jerk and snap constraints.
  • Circular trajectories: A 5 m-radius circular trajectory at 7 m s−1 required centripetal acceleration close to 10 m s−2 and was tracked near the UAV’s physical limits.The SE(3) controller achieved average position errors of 0.5 m and 0.1 m for the two circular-trajectory cases.
  • Estimator disturbances: For a 5 m jump in estimated position, MPC smoothly minimized control error while satisfying dynamics constraints.The comparison evaluated controller reactions to localization-system malfunctions or abrupt reference changes.

9 Pushing the frontiers of UAV research

The platform has supported diverse UAV research and deployment scenarios, including swarms, sensing, aggressive control, indoor inspection, and robotics competitions. These applications shaped a universal architecture validated beyond laboratory conditions.

  • Platform scope: The system was developed through extensive applied research and is presented as a universal platform for varied sensor configurations and application scenarios.The publication focuses on the underlying system rather than any single application.
  • Demonstrations: Figures document simulated and real circular tracking, aggressive maneuvers, mutual localization, and hardware deployment on Tarot 650 and custom airframes.These examples connect the platform’s simulation, control, sensing, and hardware capabilities.
  • Multi-UAV research: Real-world experiments used the platform for UAV detection, mutual localization, cooperative surveillance, complex maneuvers, and multi-UAV load carrying.Mutual localization used the open-source UVDAR system with UV blinkers and onboard filtered cameras.
  • Aggressive aerial manipulation: The platform supported aggressive control and optimal fire-extinguishing trajectories, including almost-free-fall water-dispatch maneuvers.The SE(3) geometric feedback was reported as crucial for verifying trajectory feasibility.
  • Indoor inspection: Indoor inspection used a UAV equipped with 3D LiDAR to capture detailed imagery of hard-to-reach points in historical buildings.The platform was also used for ionizing-radiation mapping and radio-source localization.

9.5 UAV swarms and formations

The system was validated in UAV swarms, GNSS-denied exploration, international competitions, and education. These deployments demonstrate its use across coordinated autonomy, subterranean navigation, and hands-on multi-robot research.

  • 9.5 UAV swarms and formations: Flocking experiments evaluated UAV formations with GNSS localization and onboard sensing only in forest environments.The work also addresses coordination in GNSS-denied navigation while minimizing object-estimation error.
  • 9.6 MBZIRC 2017 competition: In MBZIRC 2017, the team won first place for autonomous ferrous-object gathering and second place overall after achieving the fastest autonomous landing on a moving car.Both tasks were tackled using the proposed system.
  • 9.7 The DARPA Subterranean (SubT) challenge: During the DARPA SubT Tunnel Circuit, autonomous UAVs explored underground mines, localized objects, and returned safely, earning first among self-funded teams and third overall.The team reported exploring farther into the tunnels than other teams to its knowledge.
  • 9.7 The DARPA Subterranean (SubT) challenge: In the DARPA SubT Urban Circuit, UAVs explored 2867 m3 and navigated up to 100 m in 200 s through an unknown reactor-building environment.The team again placed first among self-funded teams and third overall.
  • 9.8 MBZIRC 2020 competition: All MBZIRC 2020 tasks were solved using the system, including autonomous wall building, firefighting, ball catching, fire-blanket deployment, and balloon popping.The competition helped consolidate many of the platform’s functionalities.
  • 9.9 IEEE RAS Summer School on Multi-robot Systems: More than 70 international students used the platform to solve a multi-UAV Dubins traveling-salesman problem with neighborhoods and test solutions outdoors.The system served as an educational tool during the 2019 IEEE RAS summer school.

10 CONCLUSIONS

The presented multirotor UAV control and estimation system emphasizes realistic simulation, real-world experimentation, and deployment beyond laboratory conditions. Its architecture supports agile maneuvers and safe flight with noisy state estimates.

  • The system was developed for realistic simulations and real-world experiments.
  • Its architecture enables reliable UAV deployment outside laboratory conditions using only onboard sensors.
  • The control pipeline supports both fast, agile maneuvers and safe flight with noisy state estimates.

11 Declarations

The declarations report compliance with applicable guidelines, author contributions and funding, open-source software availability, and no conflicts of interest.

  • The authors state that all applicable institutional and national guidelines were followed.
  • Informed consent was obtained from all co-authors.
  • The authors describe contributions spanning system development, controller design, state estimation, software, hardware, guidance, and manuscript preparation.
  • The work was supported by CTU, Czech Science Foundation, Ministry of Education of the Czech Republic, and European Union funding.
  • The authors declare no conflict of interest.
  • The software is available as open source, with additional multimedia materials provided online.
Loading 2008.08050v6…