DEMO / RECORDING
Overview & result
A MuJoCo quadrotor uses camera-derived symbolic state for Jev tactical judgments, while ordinary code handles perception, 500 Hz control and a 50 Hz safety reflex. The repo includes an original course-run video and a no-Jev ablation.
From the original post
everything time-critical stays in ordinary code.
How Jev fits in the loop
- Convert onboard depth and segmentation into range sectors, obstacle height and target visibility.
- Ask Jev for a bounded maneuver, risk score and target-lost judgment only when the scene changes.
- Let guidance and hard reflex code validate or override the proposal before controlling the drone.
How to reproduce
- Clone the public repository, install its documented MuJoCo/Python dependencies and configure a TypeSafe key.
- Run the Jev flight and the `--no-jev` ablation with the documented seeds.
- Inspect the telemetry video, trace and safety reflex overrides rather than treating a successful clip as a general flight benchmark.
Why this build matters
Separates low-rate model judgment from time-critical safety control in a real agent loop.
Limitations
- Jev sees symbolic camera-derived state, not camera pixels, and is advisory at about 2.5 Hz.
- A deterministic 50 Hz reflex can veto Jev; the author-run ablation uses a specific five-station simulated course.