Community · E2 · artifact verified
Choose drone tactics from camera-derived state
A simulated quadrotor uses Jev for low-frequency tactical judgments while classical vision, flight control, and safety reflexes remain in code.
01 · Role in the system
What Jev does here
Jev reads symbolic scene features derived from onboard depth and segmentation buffers. It chooses a maneuver, scores risk, and estimates whether a target is truly lost. A 50 Hz reflex layer can veto those judgments, while a 500 Hz geometric controller owns flight dynamics; Jev never receives pixels or direct motor control.
Layered authority: the model advises, reflexes veto
The architecture is the claim. Jev reads symbolic features derived from depth and segmentation buffers and answers tactical questions — maneuver choice, risk score, whether a target is genuinely lost — while a 50 Hz reflex layer can veto any judgment and a 500 Hz geometric controller owns flight dynamics. The model never sees pixels and never touches a motor.
This is the opposite shape from end-to-end learned flight: perception, safety, and tactics are separate code paths with explicit authority, and the model's judgment sits exactly where latency tolerance is highest. As a template for physical systems, that separation matters more than any single course result.
An honest pair of results
The 77.5-meter obstacle-course distance is one simulation run in MuJoCo, and the record pairs it with the less flattering finding that the matched multi-seed arena comparison showed no advantage. Keeping both numbers visible is what evidence hygiene looks like in this directory.
What would upgrade this record is straightforward: repeated runs across seeds with the same instrumentation. The repository is public and the harness logs its judgments, so the missing win-rate is a matter of compute and time, not access.
02 · Control boundary
Where Jev sits
Classical perception to Choice, Score, and Noul judgments; faster deterministic loops retain safety authority.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- The strongest course result is one simulation run and does not establish a repeatable improvement.
- Perception and safety are separate code paths; the model neither sees camera pixels nor controls motors directly.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- RomanSlack ↗Community · github · public · checked 2026-09-18