jev-drone
jev-drone (github.
Overview
jev-drone (github.com/RomanSlack/jev-drone) simulates a camera-only quadrotor in MuJoCo. Classical CV at fifteen hertz builds symbolic range sectors, obstacle height, and target bearing; TypeSafe Jev at about two and a half hertz advises maneuver Choice (hold_course, gap_left, gap_right, climb, brake, reacquire), risk Score, and target_truly_lost Noul when the scene fingerprint changes. A five hundred hertz geometric controller and fifty hertz safety reflex always own the sticks; code vetoes unsafe climbs. README ablation reports the no-Jev baseline never passes station two while Jev engaged clears the full course in the author's single documented run.
Problem: Pure geometric planners stall on maneuvers like climbing when gaps do not exist, yet vision models are too slow and opaque for 500 Hz control.
Built for: Robotics curious developers evaluating advisory Jev loops on top of classical sim controllers.
First indexed on Jev Directory: 2026-09-17
Creator and team
- Name
- Roman Slack
- Handle
- @RomanSlack
How Jev is used
- Role in the product flow
- Low-rate tactical advisor over symbolic perception, not pixel input
- Primitives
- ChoiceScoreNoul
- State in
- Compact JSON scene: forward range sectors, obstruction height, top-edge visibility, target bearing, and fingerprint hash from depth plus segmentation buffers.
- Decision out
- maneuver Choice, risk Score, and target_truly_lost Noul probabilities consumed by guidance with hard reflex overrides.
- Eye pipeline segments depth at fifteen hertz without ground truth cheats
- Fingerprint unchanged scenes reuse last Jev judgment
- Batch three questions in one Jev call when tactical context shifts
- Guidance merges Jev advice with geometric controller and reflex vetoes
Roman's drone is the clean split brain story: perception and control stay in code you can unit test; Jev only answers small meaning questions when the symbolic scene changes. Without climb in the action vocabulary the greedy baseline dies at the low beam; Jev supplies the maneuver label while reflexes prevent suicide climbs. README is candid that early arenas showed no advantage and that the published Jev column is a single run, not a seed average. Try the Vercel demo for visuals; there is no hosted X clip in this batch, so the thick page leans on README figures and the interactive site.
Sourced performance claims
- README ablation table: baseline without Jev stops near 17.7 m; Jev engaged run reaches 77.5 m full course with zero collisions in the cited run.Source: github.com/RomanSlack/jev-drone README
- Typical sixty five second flight issues about one hundred ten Jev calls at 0.11 s median latency per README.Source: github.com/RomanSlack/jev-drone README
- Public GitHub repo RomanSlack/jev-drone had about one hundred thirty one stars when this listing was drafted.Source: GitHub star count September 2026
Features and stack
Features
- MuJoCo Skydio-style airframe with five station course
- Advisory Jev at ~2.5 Hz with fingerprint caching
- 500 Hz controller plus 50 Hz safety reflex overrides
- Web demo at jev-drone.vercel.app
Stack
- Python
- MuJoCo
- Classical CV
- TypeSafe Jev
- Vercel demo frontend
Pricing: Open source sim; live Jev calls need your TypeSafe API key during runs.
Links
FAQ
- Does Jev see camera pixels?
- No. README states Jev reads JSON scene summaries built by classical CV, not raw images.
- Can Jev override safety reflexes?
- No. Fifty hertz reflexes and geometric control veto unsafe maneuvers even if Jev proposes climb.
- Are the ablation numbers averaged?
- README warns the Jev column is one sixty five second run and earlier arenas showed variance. Treat claims as structural, not leaderboard guarantees.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.