Skip to content
AIBRAINWORX / THE INDEPENDENT AI LABRESEARCH · HARDWARE · OPEN INTELLIGENCE
Aviation & Drones / Robotics & Embodied / Developer & Agent Tools / Robotics

PX4 Autopilot v1.17 Brings Mature Zenoh Middleware and Expanded ROS 2 Controls

PX4 v1.17 adds Altitude Cruise flight mode, improved fixed-wing navigation-loss recovery, and tighter rmw_zenoh and ROS 2 fixed-wing interfaces.

Listen to this articleGoogle Voice
0:00
PX4 Autopilot emblem on dark slate background
Official PX4 project emblem styled cleanly for AiBrainWorX and OpenSourceCoder.
Event / context: 2026-05-13. Source check: October 28, 2026.

The Dronecode PX4 project shipped PX4-Autopilot v1.17.0, a major stable release advancing modular robotics middleware, fixed-wing and VTOL companion-computer control, inertial navigation drivers, and on-device neural-network flight control research. Building atop the v1.16 release branch, v1.17 packages dozens of core safety, simulation, and MAVLink improvements tailored for commercial autonomy teams and academic drone labs alike.

Experimental On-Device Neural Network Control via TFLite Micro

One of the headline additions for A.I. flight researchers is an initial neural-network multicopter control test path powered by TensorFlow Lite Micro integrated directly inside firmware. Through the `mc_neural_network_control` module, developers can train reinforcement-learning policies in external physics simulators such as Aerial Gym, export quantized `.tflite` weights, and evaluate closed-loop neural control on bench hardware or controlled test vehicles.

PX4 maintainers explicitly emphasize that the TFLite Micro flight path is strictly intended for research and bench testing rather than replacing the battle-tested deterministic PID flight controller in production operations. Still, offering a standardized, reproducible onboard inference hook lowers the friction of moving reinforcement-learning policies off offboard GPUs onto flight microcontrollers.

Maturing Zenoh Middleware and Clean ROS 2 Control Interfaces

Companion-computer integration takes a major step forward in v1.17 as PX4's in-tree Zenoh middleware reaches full `rmw_zenoh` compatibility—adding CDRv1 serialization, ROS 2 graph liveliness announcements, domain ID configuration, Zenoh CLI diagnostics, and auto-generated message mapping from `dds_topics.yaml`. Zenoh ships enabled in default firmware for FMU-v6xRT boards and as dedicated `_zenoh` build targets for FMU-v6x and SITL simulation, offering lighter embedded overhead and cleaner multi-hop wireless traversal than raw DDS over contested Wi-Fi or mesh links.

To tighten offboard trajectory scripting, fixed-wing aircraft and VTOL planes in forward flight gain `FwLateralLongitudinalSetpointType` in the PX4 ROS 2 Control Interface, alongside `RoverSetpointTypes` providing type-safe setpoint combinations (position, speed, throttle, attitude, steering) for uncrewed ground rovers.

New Flight Modes, Failsafes, INS Drivers, and Gazebo Jetty

Pilots and fleet operators also pick up practical safety and manual-flight refinements: multicopters gain a new Altitude Cruise mode that preserves tilt and heading when sticks return to center instead of braking to a full stop, while Fixed Wing Takeoff mode now maintains a level-wing climb and can loiter around takeoff coordinates if navigation locks drop.

Rounding out hardware and simulation support, v1.17 introduces INS drivers for MicroStrain, sbgECom, and EULER-NAV units, adds Septentrio GNSS resilience reporting and barometer-vs-GNSS height auto-calibration, and brings simulation compatibility forward with Gazebo Jetty and Ackermann simulator-in-hardware support.

Closing the Gap Between Edge A.I. Policies and Real-World Avionics

Historically, aerospace researchers training neural-network flight controllers inside GPU-accelerated physics simulators faced a brittle engineering chasm when transferring policies onto physical airframes: translating Python research loops into deterministic, lock-free C++ modules running safely alongside state estimators, battery monitors, and manual-override RC switches. By embedding TensorFlow Lite Micro directly inside PX4 v1.17's controlled experimental `mc_neural_network_control` module, teams can test quantized `.tflite` control graphs natively on flight hardware while retaining deterministic firmware watchdogs and manual safety pilot switchover.

Paired with mature `rmw_zenoh` middleware for low-overhead wireless links between companion computers and flight controllers, plus explicit ROS 2 fixed-wing and rover setpoint APIs, PX4 v1.17 strengthens its role as a clean worldwide foundation linking academic simulation, commercial autonomy software, and open-source autopilot silicon without breaking deterministic production flight stacks.

Primary-source release report from PX4-Autopilot GitHub; verified November 17, 2026.

Follow your curiosity.

Pass the signal.