ModalAI Forum

Re: Starling 2 Max no GPS data
Re: Poor GPS Fix

Dear ModalAI Support Team,

We are encountering similar GPS issues with the Starling 2 platform. After reviewing various forum posts and discussions, we have found multiple proposed solutions; however, the information appears fragmented and in some cases outdated.

Could you please provide an updated summary of the current recommended short-term mitigations and any permanent fixes? Additionally, is there an official channel (e.g., release notes, issue tracker, or mailing list) where ongoing updates regarding this issue are communicated?

Thank you for your assistance.

Re: Apriltag relocalization not relocalizing?

Hello, I have the following issue similar to one above about apriltag relocalization with fixed frame:

Platform / Version:
Starling 2 (C27)
voxl-suite v1.5.0
Front-facing camera used for VIO
Down-facing camera used for AprilTag detection

Issue
AprilTag relocalization works correctly when the tag detector runs on the front-facing camera, but consistently fails when running on the down-facing camera, even though:

• The down-facing camera intrinsics are correct
• The camera→body extrinsics in /etc/modalai/extrinsics.conf are correct
• R_tag_to_fixed is correctly set for a flat ground tag R_tag_to_fixed= [[0, -1, 0], [1, 0, 0], [0, 0, 1]]
• The tag is physically flat and the vehicle is level
• With the down-facing camera, voxl-vision-hub repeatedly reports: "WARNING, apriltag roll/pitch out of bounds" does not relocalize.

Confirmation
Using the same tag, same fixed-frame configuration, and same environment:

• Detection from front camera → relocalization succeeds.
• Detection from down-facing camera → relocalization rejected.
• Swapping only the camera input is sufficient to reproduce the failure.

Conclusion
This strongly suggests that the AprilTag relocalization path in voxl-vision-hub is composing R_tag_to_cam → R_cam_to_imu/body using the wrong camera extrinsics (appears to default to the front/tracking camera), rather than the extrinsics corresponding to the camera that actually produced the detection. As a result, the inferred R_tag_to_local ∘ R_fixed_to_tag contains spurious roll/pitch and is rejected by the ±10° gate.

Could you please solve the issue in new SDK? (You might add additional parameters in /etc/modalai/voxl-vision-hub.conf to select fixed frame pipe)

Best,
Kağan

Dear all,

I would like to ask you a couple of questions about reliably simulating the Starling-2 drone in a robotic simulator (preferably Gazebo).

• Do you have/provide an SDF of Starling-2 for Gazebo?
• Do you have some documentation for simulation in general that is similar to this (https://developer.parrot.com/docs/sphinx/index.html)?
• Do you have some system identification results for the motor and propeller couple (speed vs thrust)?
• I know the autopilot part (PX4) can be run as software in the loop (https://docs.px4.io/main/en/simulation/), but I wonder if the VOXL SDK can be run in an Ubuntu desktop computer (as a ROS node or non-ROS)?

Thank you in advance.
Best regards,
Kağan

@Moderator Thank you for your reply but I am able reach only the files of latest version of Starling from the datasheet and I need the CAD files of the one below if you have it:

Thank you in advance. Best,
Kağan

Hello,

I am working with the initial version of the Starling drone (V1) in our lab at EPFL. Do you have the CAD files for this version? I need the frame files, in particular.

Best,
Kağan