CERTIFIED

< Droneforge > Interested in testing the drone-ai state of the art, without having to spend 100 hours setting up your drone? Check out Nimbus - our product that wirelessly enables autonomy to off-the-shelf drones, allowing you to spend time on what matters. Flying.

< Real-World Flight Tests > The system was tested on a real quadrotor in three environments of increasing clutter A user guided the drone from a start to a goal position using only the sketching interface Sketchplan achieved 100% task success in low/medium clutter Even in the

< Collecting the Dataset > Collecting thousands of real-world drone flights with matching sketches is impractical Instead, the team built a dataset entirely in simulation using photorealistic 3D Gaussian Splatting (3DGS) scenes They generated 32,000+ smooth 3D paths, then had

< The Solution: A Modular, Diffusion-Based Planner > Sketchplan uses a two-part architecture: > SketchAdapter: A tiny model trained on a small, human-labeled dataset (872 sketches) that learned to map a human's freehand drawing to a "perfect" 2D projection of a 3D path >

< The Challenge: Intuitive Robot Instruction > Telling a robot exactly where to go is hard Language is often too vague, and teleoperation is too low-level Sketches offer a perfect middle ground because its: > embedded with fine-grained spatial intent > quick and natural for

< SketchPlan: Fly a Drone by Drawing Its Path > Researchers just dropped Sketchplan, a new path planner that lets you control a drone by simply drawing a path on a screen It's a diffusion-based model that interprets your 2D sketch over a live depth image and generates a smooth

introducing our click-and-nav agent! made with 6 lines of code and deployed on $100 hardware

EdgeTAM now supported by @thedroneforge what should we do with this model and an autonomous drone?

https://t.co/0r7kEQ5DvZ

full self driving for drones deployed today on a $100 drone with monocular vision and no external localization sensors

my drone doesn't like to ram into something and he is scared of people!

https://t.co/ZrYHqiQwdg

|￣￣￣￣￣￣￣￣￣￣￣￣￣￣| | i <3 @thedroneforge | |＿＿＿＿＿＿＿＿＿＿＿＿＿＿| \ (•◡•) / \ / —— | | |_ |_

we turned a $100 stupid drone into a state-of-the-art fully autonomous drone > constructed in 3 minutes > only 7 lines of code > 10 km max range thx @miamishor for stopping by :)

It's been feeling like the Manhattan project at the office lately and I crave more.

< Intelligent Drone Agents > At Droneforge, we put state-of-the-art drone intelligence at your fingertips Allowing you to program and deploy customizable Drone Agents for real-world tasks Checkout us out! https://t.co/1wqAxBvT6j 🧵6/6

< Compared to previous SOTA > > First to reach champion-level speed > First to run entirely onboard > First to handle visual ambiguity using decoded progress tracking > First to close the visual sim-to-real gap with low-quality masks 🧵5/6

< Real-world performance > > 21 m/s top speed > 6 g acceleration > Executes inverted loops, split-S, ladders > Survives poor segmentation and battery drop-off > Adapts to 30 % thrust loss by re-estimating max motor RPM mid-flight.... ...and keeps racing 🧵4/6

< END-T0-END! > The beauty of the whole thing is that it maps segmentation masks + IMU + motor RPMs DIRECTLY to motor commands. There is no need for PID or inner loop The policy runs fully onboard (Jetson Orin NX) at 90 Hz. End-to-end, no external aid, no calibration, no