GitHub · by Roman Slack

jev-drone

A simulated drone that asks Jev for a judgment 2.5 times a second.

Roman Slack’s drone project. The repository describes a camera-only autonomous drone in the MuJoCo simulator, with Jev in the control loop at 2.5 Hz.

The airframe is the Skydio X2 from MuJoCo Menagerie, a real four-rotor aircraft flown by a geometric controller, and nothing about the flight is scripted. The stack is layered by rate: a 500 Hz geometric controller, a 50 Hz guidance and safety reflex that always owns safety, a 15 Hz step that turns the camera’s depth and segmentation buffers into a compact symbolic scene, and Jev at about 2.5 Hz for tactical judgment, advisory only.

Jev is not a vision model and cannot be the perception layer. Classical computer vision does that work and hands over five forward range sectors, the height of whatever is blocking the path, whether its top edge is visible and where the target is. Jev answers three questions about that scene in one call: which manoeuvre (hold course, gap left, gap right, climb, brake or reacquire), how risky the situation is, and whether the target is genuinely lost or just briefly occluded.

Code decides when to ask, which is what keeps the cost down. On an open corridor with the target in view there is nothing to decide and no call is made, and scenes are fingerprinted so an unchanged situation reuses the last judgment — a typical 65-second flight costs about 110 calls. Code also keeps the veto: the 50 Hz reflex overrides any judgment when something is close, and Jev’s proposal to climb is refused unless the obstruction’s measured top edge is within the aircraft’s climb ceiling.