Post
252
Hey !
I'm working on a small-scale multi-drone control system and I'm looking for an open-source VLM that can run in real time on a Jetson Orin. If anyone knows a model or is personally interested in this kind of edge robotics problem, I'd love pointers.
What I'm trying to solve :
I have 4 simultaneous video streams coming from four drones (grayscale, 320×320 ). I can feed the model either:
• a 2×2 mosaic frame, or
• 4 separate frames as a batch.
Along with this, I provide a short text instruction describing the mission state.
What I need from the model :
A single structured JSON command representing the next action for the swarm controller. Something like (not decided yet):
So I need a VLM that can:
• handle multi-image or mosaic image input
• run efficiently on a Jetson Orin (ideally INT4/INT8 friendly, TensorRT-compatible)
• generate stable JSON outputs based on visual + textual context
I would really appreciate suggestions, or even just thoughts on what architectures make sense here.
Models like openbmb/MiniCPM-V-4_5, dustnehowl/nanoVLM, and Qwen/Qwen3-VL-8B-Instruct look promising, but I'm still exploring what’s actually viable on-device.
Happy to share benchmarks or test anything people want to throw at this problem. The multi-drone video + action JSON setup is niche but potentially useful to others building edge-deployed agents.
I'm working on a small-scale multi-drone control system and I'm looking for an open-source VLM that can run in real time on a Jetson Orin. If anyone knows a model or is personally interested in this kind of edge robotics problem, I'd love pointers.
What I'm trying to solve :
I have 4 simultaneous video streams coming from four drones (grayscale, 320×320 ). I can feed the model either:
• a 2×2 mosaic frame, or
• 4 separate frames as a batch.
Along with this, I provide a short text instruction describing the mission state.
What I need from the model :
A single structured JSON command representing the next action for the swarm controller. Something like (not decided yet):
{
"action": "move_forward",
"confidence": 0.87,
"reason": "front corridor detected, no obstacles in drone_2 and drone_4 views"
}So I need a VLM that can:
• handle multi-image or mosaic image input
• run efficiently on a Jetson Orin (ideally INT4/INT8 friendly, TensorRT-compatible)
• generate stable JSON outputs based on visual + textual context
I would really appreciate suggestions, or even just thoughts on what architectures make sense here.
Models like openbmb/MiniCPM-V-4_5, dustnehowl/nanoVLM, and Qwen/Qwen3-VL-8B-Instruct look promising, but I'm still exploring what’s actually viable on-device.
Happy to share benchmarks or test anything people want to throw at this problem. The multi-drone video + action JSON setup is niche but potentially useful to others building edge-deployed agents.
