jev-drone

# jev-drone

An autonomous quadrotor flies a five-station obstacle course in MuJoCo using
**only its onboard camera**. A small judgment model ([TypeSafe](https://typesafe.ai)'s
Jev) sits at ~2.5 Hz and decides *what the situation means*; everything
time-critical stays in ordinary code.



The drone is the Skydio X2 from [MuJoCo Menagerie](https://github.com/google-deepmind/mujoco_menagerie)
- a real airframe with four rotors, flown by a geometric controller. Nothing
about the flight is scripted.

## The idea

Jev is **not** a vision model. It takes JSON and returns typed answers with
probabilities. So it cannot be the perception layer, and it cannot run at
control rate. What it can do is answer small questions about a situation that
ordinary code finds hard to phrase.

```
 500 Hz   geometric controller       flight.Pilot    thrust-priority mixing, slew-limited
  50 Hz   guidance + safety reflex   run.Guidance    ALWAYS owns safety
  15 Hz   camera -> symbolic scene   flight.Eye      depth + segmentation, no ground truth
~2.5 Hz   tactical judgment          tactics.py      Jev, advisory only
```

Classical CV turns the depth and segmentation buffers into a compact scene:
five forward range sectors, the height of whatever is blocking the path,
whether its top edge is even visible, and where the target is. Jev reads that
and answers three questions in one call:

| question | type | options |
|---|---|---|
| `maneuver` | **Choice** | hold_course / gap_left / gap_right / climb / brake / reacquire |
| `risk` | **Score** | clear and open -> tight -> about to hit something |
| `target_truly_lost` | **Noul** | genuinely lost, or just briefly occluded? |

Code decides *when* to ask. On an open corridor with the target in view there
is nothing to decide, so no call is made. Scene