We are currently evaluating off-the-shelf development platforms for an upcoming computer vision project. We are looking closely at the Starling 2 Max (and/or custom VOXL 2-based platforms) and would love to verify feasibility against our core requirements.

Project Overview & Requirements

Our pipeline runs a custom C++ algorithm combined with deep learning model inference onboard the aircraft, alongside manual flight control.

Here are our primary hardware and software requirements:

1 High-Res Video & Uncompressed/Raw Access:
◦ Resolution: Minimum 4K native video resolution (8K capability preferred).
◦ Data Pipeline: We require direct, low-latency access to uncompressed / raw frame buffers (or near to it)
◦ Optics (optional): Ability to support continuous optical zoom (at least 4x).

2 Onboard AI Compute & Performance:
◦ Our models are currently built/optimized for PyTorch and NVIDIA TensorRT.
◦ What real-world FPS / latency can we expect for standard object detection models (e.g., YOLO variants) ?

3 Flight Range, Autonomy & Control:
◦ Full manual flight control capabilities for the pilot.

Key Questions for the ModalAI Team

1. Camera & Zoom Integration: * The Starling 2 Max standard setup features IMX412 (12MP) sensors. Can we stream full 4K uncompressed raw frames via MIPI into custom C++ code seamlessly?
2. Custom C++ & Library Support: How simple is it to compile and run external static/dynamic C++ libraries directly within the VOXL SDK runtime environment?
3. Companion Compute Option: If our inference workloads heavily rely on NVIDIA TensorRT, is it feasible to mount a secondary compute payload (e.g., NVIDIA Jetson Orin Nano/NX) onto the Starling 2 Max frame using VOXL 2 Ethernet/USB3 expander boards, and what would be the impact on payload capacity (~500g limit) and flight time?

Looking forward to your guidance

Thank you so much for your attention and participation.

Sebastian