* refactor(rover): run each rover controller on its own trigger topic The rover modules ran the whole controller cascade from the gyro callback, so the position and speed controllers ran several times per EKF output on stale estimates. Each controller is now a work item triggered by the topic it consumes (local position, attitude, angular velocity, steering setpoint), and the module only generates setpoints and enables the controllers the active control mode needs. All work items share the rate_ctrl queue, so enabling and resetting never overlaps a controller run. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * fix(rover): register controller callbacks top-down Position and speed control share the vehicle_local_position trigger and uORB schedules callbacks in registration order. Speed was registered first, so it always consumed the previous cycle's position-control setpoint, adding one EKF period of latency. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * docs(rover): move controller update descriptions back to their functions The init() and setEnabled() declarations were inserted between the update*Control() doc comments and their functions, leaving two stacked @brief blocks. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * fix(rover): widen the dt clamp of the EKF-triggered controllers Attitude and speed control now run on vehicle_attitude and vehicle_local_position instead of the gyro. Their period is the EKF2 output period, nominally exactly the old 10 ms upper bound, so jitter or a larger EKF2_PREDICT_US clipped dt and slowed the slew-rate limiters and PID integrators below real time. Allow up to 50 ms. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * docs(rover): correct the attitude controller dt comment vehicle_attitude is published on every EKF2 output predictor update at IMU_INTEG_RATE, not at the EKF2_PREDICT_US filter update rate. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * fix(rover): restart the controller dt when a controller is re-enabled A controller keeps its last run timestamp while it is disabled, so the first run after a mode switch computed dt over the whole disabled time and hit the upper clamp. With the speed and attitude clamps at 50 ms that step was six EKF periods long for the slew limiters and integrators. Restart the timestamp on enabling, so the first dt is the lower bound. SIH rover (10045), position -> manual -> position while driving, speed PID integral after the first re-enabled run: dt 50 ms before, 1 ms after (normal runs 8 ms). Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * fix(rover): pace the actuator controllers by the gyro The actuator controllers ran only when a steering setpoint arrived. A mode change resets the rate controller's setpoint to NaN, so the next cycle published no steering setpoint and therefore no actuator setpoint. SIH lockstep waits for an actuator output every step and froze there: switching to stabilized or offboard velocity stalled the simulation. Trigger the actuator controllers on vehicle_angular_velocity, registered after the rate controller, so they publish every cycle with the steering setpoint of the same cycle. SIH rover (10045), manual/acro/stab/position/offboard velocity cycle, 3 runs each: before all 3 runs froze (offboard switch in all, stab switch in one), after no lockstep timeouts, actuator outputs every cycle (max gap 8 ms) and >= 99.7 % of them use the steering setpoint of the same cycle. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * perf(rover): only wake the enabled controllers All controllers stayed registered on their trigger topics, so in every mode and while disarmed the disabled ones were scheduled at up to gyro rate just to return. Register a controller's callback when it gets enabled and unregister it when it gets disabled. A newly enabled controller runs once right away, because it registered after the trigger of the current cycle fired; without that the lockstep simulation misses the cycle's actuator output and stalls. When an upper controller is enabled while the lower one sharing its trigger topic keeps running (position/speed on vehicle_local_position, rate/actuator on vehicle_angular_velocity), the lower one is re-registered behind it to keep the cascade order. SIH rover (10045), 3 runs each: - work_queue status: disarmed only the module runs, manual adds the actuator controller, acro rate + actuator, stab + attitude, position all five, offboard velocity all but position; no lockstep timeouts. - Value-based order check (servo command vs steering setpoint, speed controller setpoint vs position controller setpoint): every consumer uses the setpoint of the same cycle after manual -> acro, manual -> offboard velocity and offboard velocity -> Go-To. With the re-registration disabled the actuator controller used the previous cycle's steering setpoint after acro / offboard, and the speed controller the previous cycle's speed setpoint in Go-To (153 of 153). - Re-enable probe: first dt after position -> manual -> position is still 1-2 ms. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * fix(rover): assign the rover control workitems to nav_and_controllers * fix(rover): retry a failed controller callback registration setEnabled() marked a controller enabled even when registering its trigger callback failed. The next call then saw no change and never retried, so the controller ran once from ScheduleNow() and then stayed idle with a single error. A controller now only counts as enabled once its callback is registered, so the next module cycle retries. A failed re-registration for the cascade order clears it the same way. The modules report a failure once instead of every cycle. SIH rover (10045), attitude controller registration forced to fail for the first 1500 attempts: the controller stays off while it fails and starts at 250 Hz on the first successful retry, one error per failure episode. Mode cycle without the fault unchanged. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * refactor(rover): share the controller trigger logic in a base class All 15 rover controllers carried the same copy of init(), setEnabled(), enabled(), reregisterTrigger() and the enabled check in Run(). Copies of the code that decides when a controller runs and in which order can drift apart, so RoverControllerWorkItem in lib/rover_control now owns the trigger subscription, the work queue and the enable state. A controller only implements its step and, if it has a time step, the restart on enabling. The trigger only wakes the controller, so the controllers read their inputs through plain subscriptions again. The actuator controllers no longer drain the gyro topic: callbacks are scheduled on new publications, not on unread data. A functional test covers the contract the modules rely on: no runs while disabled, one run right away on enabling, a run on every trigger, registration order for controllers sharing a trigger and re-registering to restore the cascade order. SIH rover (10045), manual/acro/stab/position cycle: each mode wakes the same controllers at the same rates as before, actuator setpoints at 250 Hz, no errors. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> * fix(rover): drop the protected destructor of the controller base class clang-tidy (cppcoreguidelines-virtual-class-destructor) rejects a destructor that is both protected and virtual. WorkItem already has a public virtual destructor, so the implicit one is enough. Assisted-by: Claude:claude-opus-5-5 Signed-off-by: Michael Fritsche <frimi999@gmail.com> --------- Signed-off-by: Michael Fritsche <frimi999@gmail.com>
The autopilot stack the industry builds on.
About
PX4 is an open-source autopilot stack for drones and unmanned vehicles. It supports multirotors, fixed-wing, VTOL, rovers, and many more experimental platforms from racing quads to industrial survey aircraft. It runs on NuttX, Linux, and macOS. Licensed under BSD 3-Clause.
Why PX4
Modular architecture. PX4 is built around uORB, a DDS-compatible publish/subscribe middleware. Modules are fully parallelized and thread safe. You can build custom configurations and trim what you don't need.
Wide hardware support. PX4 runs on a wide range of autopilot boards and supports an extensive set of sensors, telemetry radios, and actuators through the Pixhawk ecosystem.
Developer friendly. First-class support for MAVLink and DDS / ROS 2 integration. Comprehensive SITL simulation, hardware-in-the-loop testing, and log analysis tools. An active developer community on Discord and the weekly dev call.
Vendor neutral governance. PX4 is hosted under the Dronecode Foundation, part of the Linux Foundation. Business-friendly BSD-3 license. No single vendor controls the roadmap.
Supported Vehicles
|
Multicopter |
Fixed Wing |
VTOL |
Rover |
…and many more: helicopters, autogyros, airships, submarines, boats, and other experimental platforms. These frames have basic support but are not part of the regular flight-test program. See the full airframe reference.
Try PX4
Run PX4 in simulation with a single command. No build tools, no dependencies beyond Docker:
docker run --rm -it -p 14550:14550/udp px4io/px4-sitl:latest
Open QGroundControl and fly. See PX4 Simulation Quickstart for more options.
Build from Source
git clone https://github.com/PX4/PX4-Autopilot.git --recursive
cd PX4-Autopilot
make px4_sitl
Note
See the Development Guide for toolchain setup and build options.
Documentation & Resources
| Resource | Description |
|---|---|
| User Guide | Build, configure, and fly with PX4 |
| Developer Guide | Modify the flight stack, add peripherals, port to new hardware |
| Airframe Reference | Full list of supported frames |
| Autopilot Hardware | Compatible flight controllers |
| Release Notes | What's new in each release |
| Contribution Guide | How to contribute to PX4 |
Community
- Weekly Dev Call — open to all developers (Dronecode calendar)
- Discord — Join the Dronecode server
- Discussion Forum — PX4 Discuss
- Maintainers — see
MAINTAINERS.md - Contributor Stats — LFX Insights
Contributing
We welcome contributions of all kinds — bug reports, documentation, new features, and code reviews. Please read the Contribution Guide to get started.
Citation
If you use PX4 in academic work, please cite it. BibTeX:
@software{px4_autopilot,
author = {Meier, Lorenz and {The PX4 Contributors}},
title = {{PX4 Autopilot}},
publisher = {Zenodo},
doi = {10.5281/zenodo.595432},
url = {https://px4.io}
}
The DOI above is a Zenodo concept DOI that always resolves to the latest release. For a version-pinned citation, see the Zenodo record or our CITATION.cff.
Governance
The PX4 Autopilot project is hosted by the Dronecode Foundation, a Linux Foundation Collaborative Project. Dronecode holds all PX4 trademarks and serves as the project's legal guardian, ensuring vendor-neutral stewardship — no single company owns the name or controls the roadmap. The source code is licensed under the BSD 3-Clause license, so you are free to use, modify, and distribute it in your own projects.