Categories

  • Note sure where to post? Ask questions here for direct access to the ModalAI engineering team

    Support Request Format for Best Results Feature Requests 3D Models
    2009
    Topics
    10139
    Posts

    C

    @Alex-Kushleyev @Cliff-Wong
    Subject: QVIO Trajectory Visualization Issue with 180-Degree Roll Rotation on VOXL2

    Hi,
    I'm encountering an issue with QVIO on a VOXL2 setup where the board is rotated 180 degrees around the roll axis. The drone flies well, and the VIO performance is acceptable, but the QVIO trajectory visualization (qvio and Vvhub_body_wrt) does not align with the actual drone movement.

    Problem Description:
    I moved the drone up and to the right, then returned it to its starting position. However, the QVIO visualization shows the movement rotated 180 degrees about the X-axis compared to the actual motion. Please see the attached image for reference:
    825c8aee-0b6d-4a6c-bbe2-51bee257cc33-image.png

    Extrinsics Configuration:
    Here is my current extrinsics configuration:

    "extrinsics": [{ "parent": "imu_apps", "child": "tracking", "T_child_wrt_parent": [0.05, 0.0067, -0.03], "RPY_parent_to_child": [0, 135, -90] }, { "parent": "body", "child": "imu_apps", "T_child_wrt_parent": [0.025, 0.0067, -0.000360], "RPY_parent_to_child": [180, 0, 0] }, { "parent": "body", "child": "stereo_front_l", "T_child_wrt_parent": [0.0793, -0.040, -0.0243], "RPY_parent_to_child": [0, 90, 90] }, { "parent": "body", "child": "stereo_rear_l", "T_child_wrt_parent": [-0.0297, 0.04, 0.0415], "RPY_parent_to_child": [0, -90, -90] }, { "parent": "body", "child": "tof", "T_child_wrt_parent": [0.078, 0, -0.025], "RPY_parent_to_child": [0, 90, 90] }, { "parent": "body", "child": "ground", "T_child_wrt_parent": [0, 0, 0.058], "RPY_parent_to_child": [0, 0, 0] }]

    Additional Notes:
    I tested an alternative RPY_parent_to_child for the imu_apps to tracking transformation, changing [0, 135, -90] to [180, 45, 90]. Both configurations produce the same results, suggesting they may be equivalent.
    The issue persists despite the correct flight behavior, indicating a possible mismatch in the coordinate frame or visualization pipeline.
    My version is:

    -------------------------------------------------------------------------------- system-image: 1.7.10-M0054-14.1a-perf kernel: #1 SMP PREEMPT Fri Sep 27 21:18:59 UTC 2024 4.19.125 -------------------------------------------------------------------------------- hw platform: M0054 mach.var: 1.0.0 -------------------------------------------------------------------------------- voxl-suite: 1.3.5 -------------------------------------------------------------------------------- ... voxl-portal 0.7.0 voxl-qvio-server 1.0.4 voxl-streamer 0.7.4 voxl-suite 1.3.5 voxl-vision-hub 1.8.9 voxl2-system-image 1.7.10-r0

    Questions:
    Could the 180-degree roll rotation of the VOXL2 be causing this visualization discrepancy?
    Are there specific calibration steps or configuration adjustments for QVIO to account for this rotation?
    Any suggestions for debugging the QVIO trajectory visualization?
    Any insights or recommendations would be greatly appreciated. Thank you!

  • Commonly asked questions we get from the community

    58
    Topics
    276
    Posts

    R

    Re: Support - information About Indoor mapping and localization

    Looking at the Starling v2 indoor slam for a peculiar mission. I do not currently have a Starling but wondering

    Needs
    -Drone to navigate interior of the building with minimal or no input
    -create a 3D or 2D map to identify Walls, Doorways, Windows (or be able to extract out of the 3D map)
    -BlueList

    Is this something anyone has done or looked into?

  • Questions by Product

    1926
    Topics
    10336
    Posts

    Ege YüceelE

    @Eric-Katzfey I see. Thanks for the response.

  • 398
    Topics
    1894
    Posts

    Ege YüceelE

    Hi ModalAI team. I’ve been successfully using voxl-tflite-server to run custom neural networks that take image input from the mpa/hires camera. Now, I’d like to expand to more general neural networks that also incorporate other forms of input. For example, prior waypoint information stored in memory or additional sensory data (e.g., IMU, rangefinders, or temporal data from previous frames).

    What would be the recommended roadmap to support this?

    Specifically:

    How should I modify or extend the current model helper pipeline to feed both image and non-image inputs?

    Is there support in voxl-tflite-server for multiple input tensors (e.g., combining a camera frame other forms of information)?

  • Please request review from ModalAI before building your drone or custom add-on PCB

    PCB Design Reviews System Architecture Design Reviews
    7
    Topics
    22
    Posts

    VinnyV

    Hi @Alex-Kushleyev , @QSL 170-250mm is quite long. I'm not optimistic 250mm can even work, 170 would be risky.
    We can offer our design services, but typically what that results in is a MOQ of a specific new item, such as 300-500pcs depending on the complexity (at normal MRSP which for flexes runs ~$50-80 each) and lead times would be ~14-18 weeks.
    What we would need is a .dxf outline of your intended design and we can review that for feasibility and interest. If it really is 250mm, I'd prefer to no-bid it. 200mm might be the edge here (we do not have sophisticated Sigrity tools to confirm, so going on experience here).
    M0170 is not ideally suited for back-to-back (the connectors are a bit offset from each other so the net result would be weird) but it may prove if 140mm can work + ~10 mm for the connector transition losses.
    In situations like this, I encourage the ME/Industrial Design Team to reconsider placement of sensors and main electronics relative to each other. There are just some constraints the electronics cannot meet easily. Sometimes just simply re-prioritizing the system constraints helps solve the spatial/distance problems. To extend past ~180+mm, I would encourage you use an external design firm that knows how to do mipi repeaters and high speed design. We've had one customer successfully do that, but it takes some skill and we could do it ourselves, but we just do not have the design cycles to support it this calendar year.
    Hope this helps and let us know what your thoughts are, and if pursuing us for a custom design is an option, we can reach out over email.