Blog블로그博客

Part 6 of Building and Evaluating a Drone VLA from Scratch. Instead of hoping one network learns to listen, split the job: a world model finds every target without reading the words, a 340-parameter selector lets the words pick one, and a planner flies there. Every pair goes the right way, pair success 6/10, and planning through the learned dynamics fails completely.

Part 5 of Building and Evaluating a Drone VLA from Scratch. Same small policy, same data; only the way it outputs an action changes. OpenVLA-style action tokens fix the flying, an 8-step action chunk makes it the best end-to-end model so far (pair success 3/10), and giving it more states makes it react to the words but stop stopping.

Part 4 of Building and Evaluating a Drone VLA from Scratch. A 481k-parameter RGB + text policy trained on 100 counterfactual pairs learns to fly toward the targets, but not which one: for almost every pair it flies to the same pillar whichever sentence it gets. Why, measured.

Part 3 of Building and Evaluating a Drone VLA from Scratch. 120 layouts, each flown twice from the identical start: once per sentence. How the pairs are generated, split and verified, and two bugs that only reading the output back could catch.

Part 2 of Building and Evaluating a Drone VLA from Scratch. What a 27 g simulated quadrotor can honestly borrow from a 1 kg drone’s spec sheet, the input/output contract every policy must satisfy, and how a flight is scored, all fixed before any model is trained.

Part 1 of Building and Evaluating a Drone VLA from Scratch. A CPU-only drone simulator records 34.5 frames per second, but the first recording passed every check I had written and still had the targets cut out of view.