Very cool, I'm working on a similar project but using MCP as the flight input layer. Would love to dm and discuss more about how you built it or collaborate.
I was looking into the MCP route too, and found some libraries abstracting mavlink for this use case (there’s at least one white paper documenting failure modes of LLMs trying to issue mavlink commands without an abstraction), but realized that autopilot like PX4 exists. My use case was more about autonomous flight, and it seemed better to just set waypoints and put some guards on other inputs. When paired with QGroundControl plans, all I needed to do for most flight paths was generate or update a .plan file using an LLM and other methodologies. I wasn’t super happy with the QGroundControl -> Gazebo rendering (no tie into real world terrain out of the box), but it did sort of work out of the box without too much effort!
Interesting. I've been building around that MCP abstraction and have had some early success flying in Cosys-Airsim (and Gazebo before that): https://github.com/jakedcmp/droneserver . I am starting to realize I need to break apart the MCP interface from all the other pieces of the stack for cleaner architecture but thats pending work. My flow goes like this: LLM -> MCP tools -> droneserver -> MAVSDK/MAVLink -> PX4/Ardupilot -> Cosys-Airsim Software in the Loop testing. What is novel for me is not having to learn how to fly a drone and bringing that capability into already existing technologies like PX4 autopilot. I have been attempting to code my own mission planning so I will check out QGroundControl as that might already be a solved problem and not worth building from scratch. I have also built the foundation of video streaming back from the drone so I can run video/image perception. Once I get perception working I am hoping I can build intent level autonomy where images are analyzed according to high level mission plan and potentially re-task the drone based on that. For example, the user issues simple command to fly around the property and scan for broken parts of the fence. During flight if an image of a broken fence is perceived, then the drone stops its flight and goes closer to capture additional imaging/video and document a gps location. Still hacking things together towards a real demo so the code probably wont port over well but idk. To anyone in this thread who wants to discuss further or collaborate let me know, it seems we are all working in a similar domain but from slightly different angles. Exciting area to build, I know there is big demand for solutions in this space over the next decade.
The video/image perception is why I started looking into this domain in the first place, but I haven't gotten there yet. I hadn't seen Cosys-Airsim! I bet it could work in tandem w/ QGroundControl just like Gazebo does. Might be looking into that :)
Yes, one of the axis on FPV controllers usually is not springing back to neutral position, for throttle. This may seem like a small difference it is important.
Is crowd strike like a digital twin / virtual world sandbox for autonomous drones? Do you have any additional information I could check out? Been working on autonomous drone flights but eventually need a digital world to experiment in but have yet to reach that step. Debating working with Unreal Engine or NVIDIA omniverse but unsure what the right direction is.
Unfortunately I do not plan/want to hand over control of the drone to an AI -> MCP as of know. Currently we plan to generate flight plans on a server and then each drone will request a flight plan and do its stuff.
very excited to play around. will be attempting to see if i can get character coherence between runs. the issue with the 8s limit is its hard to stitch them together if characters are not consistent. good for short form distribution but not youtube mini series or eventual movies. another comment about IP license is indeed an issue but its why i am looking towards classical works beyond their copyright dates. my goal is to eventually work from short form, to youtube to eventual short films. tools are limited in their current form but the future is promising if i get started now.
how were you able to tell this? still trying to understand what infra is better used for inference (say realtime image category matching) vs training (feeding a chatbot huge sums of data)
i really wonder how they are housing the desktop grade hardware. im so used to rack and stack servers (1U/2U), but do you really need that big of a chassis for a desktop cpu, a couple ram dimm's and some ssd? what are you're thoughts?
They have ATX cases on shelves. You can take a look at this tour of their datacenter [1] for more insight. ATX cases are visible around the 3 minute mark.
reply