Allow EKF2 to use AI-corrected IMU without disrupting the raw IMU stream

[souvik0306]

Describe problem solved by the proposed feature

I'm using ROS1 Noetic and PX4 v1.13.0

Currently, EKF2 relies on the raw vehicle_imu topic as its primary input.
My goal is to integrate an AI-corrected IMU stream into PX4’s estimation pipeline using the following flow:

1. Current Pipeline:

/mavros/imu/data → AI model → corrected IMU → new uORB topic (vehicle_imu_ai) → EKF2

2. Actual Requirement:

SITL IMU Data → AI model → Corrected IMU → EKF2

When EKF2 switches its input from vehicle_imu to vehicle_imu_ai, the raw IMU stream (/mavros/imu/data) stops publishing because EKF2 is its only active subscriber in SITL. Once it unsubscribes, the simulator pauses IMU publishing, which breaks the entire loop:

• MAVROS stops broadcasting pose.
• The AI model stops receiving input.
• EKF2 times out waiting for AI IMU data.
• The system stalls until EKF2 is switched back to the raw IMU.

A clean way to inject corrected IMU data into EKF2 without disrupting the original simulated IMU stream is currently missing.

Describe your preferred solution

Introduce an EKF2_AI module (or equivalent configuration) that:

1. Subscribes only to vehicle_imu_ai.
2. Publishes the same downstream topics as EKF2 (vehicle_attitude, vehicle_local_position, etc.).
3. Keeps a passive subscription to vehicle_imu alive in the background to maintain the SITL IMU stream, without using it for estimation.
4. Allows clean replacement of EKF2 or side-by-side operation with the default estimator.

This would make it possible to run state estimation on AI-corrected IMU data without breaking MAVROS or the simulator, while keeping the standard PX4 pipeline intact.

Describe possible alternatives

• Hard-wire EKF2 to vehicle_imu_ai: Simpler, but brittle. If the AI stream stalls, EKF2 fails and there’s no fallback. Startup ordering also becomes fragile.
• Keep current EKF2 and inject via HITL or MAVLink HIL_SENSOR: Tried, but EKF2 still uses the raw IMU and doesn’t allow easy handover.
• Source switching logic: Currently supported but causes the simulator to idle when raw IMU loses subscribers. This breaks the feedback loop.

Additional context

The main problem:

• Switching EKF2 input to the AI topic unsubscribes from the raw IMU.
• /mavros/imu/data then drops to ~1 Hz (heartbeat only).
• The AI model loses input and stops publishing.
• EKF2 reports poll timeouts and the estimation stops entirely.

As soon as I switch EKF2 back to vehicle_imu, everything recovers.

A dedicated EKF2_AI module or similar mechanism would provide a stable, modular way to integrate AI-based IMU correction without breaking the simulator pipeline.

[souvik0306]

If anyone has tried a different approach, please feel free to drop suggestions. My ultimate goal is to feed AI corrected IMU data into the EKF2 after taking from /mavros/imu/data or simply SITL IMU data.

[DronecodeBot]

This issue has been mentioned on Discussion Forum for PX4, Pixhawk, QGroundControl, MAVSDK, MAVLink. There might be relevant details there:

https://discuss.px4.io/t/ekf2-delays-with-external-imu/7769/8