User repository on GitHub
Three.js Autonomous Driving Reflex & AI Safety Brake Simulator powered by TypeSafe Jev System 1/2 Dual-Brain Architecture
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
JevPilot Reflex
Three.js autonomous-driving simulator that splits work between a System 2 slow planner and a Jev-style System 1 reflex loop at 60 Hz, with a safety brake that overrides the planner and applies full braking when time-to-collision drops below threshold.
- Stars
- 0
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Last verified
- Oct 2026
- Languages
- JavaScript
- Platforms
- Browser
- Capabilities
- Open sourceRuns locally
Highlights
- README reports a 1.5 ms per-step reflex decision running fully offline with no API key.
- The safety brake takes over below a 1.6 s time-to-collision threshold and commands full braking at -9.2 m/s².
- System 2 runs asynchronously at 1.5 Hz / 650 ms so slow reasoning never blocks the 60 FPS control loop.
- Scenes use a Tesla Model Y GLTF with Ackermann steering plus procedural city, town and highway maps.
Watch out
MIT licensed. It is a simulation with a local heuristic reflex solver, not a production driving stack, and the README is primarily written in Chinese.
Questions about this entry
Short answers drawn from the linked project or page.
No. The reflex layer runs offline at about 1.5 ms per step in the README's measurements; a remote Jev-compatible API or VLM is only an optional asynchronous System 2 path.
It monitors time-to-collision and physical limits below the planner; the README says it overrides upper-layer control and applies 100% braking once TTC falls under the critical threshold.
System 1 is a local heuristic reflex solver with multi-candidate trajectory manifold evaluation, one decision in about 1.5 ms, carrying the 60 Hz loop and chassis kinematics control. System 2 is an asynchronous slow reasoning interface compatible with a remote Jev API or VLM, running CoT at 1.5 Hz or 650 ms so it never blocks that control loop.
An industrial Tesla Model Y GLTF with Draco-compressed PBR materials, Ackermann steering kinematics, and procedural city streets, town and Interstate 8 routes. You can drive manually with WASD, steering and brakes, and the HUD adds a dock-style instrument panel, a navigation minimap, a 3-second candidate-trajectory fan and chase, first-person and bird's-eye cameras.
Its comparison table puts cloud inference at 60 calls per second for a pure end-to-end stack against 1.5 per second for the dual-brain design, about 97.5% less compute, with token rate dropping from roughly 3,600 to about 90 per second, uplink from about 80 Mbps to about 3 Mbps, and edge power from a 500W-plus server to under 12W. These are the project's own modeled figures.
Not as a driving stack. It is MIT licensed and a simulation running a local heuristic reflex solver, with a live demo hosted at manhua777.site/jev-pilot-reflex, and the README is primarily written in Chinese. The J key toggles Jev cruise, L switches to VLA+Jev mode, and the top bar switches clear, heavy rain and night weather.
The README's table compares a stopped vehicle at 120 km/h, 33.3 m/s: an end-to-end VLM at 500 ms leaves a 16.7 m reaction blind spot and 79.2 m total stopping distance, a human driver at 1,200 ms leaves 40.0 m and 102.5 m, while the Jev safety gate at 1.5 ms leaves 0.05 m and 62.8 m. Every 100 ms of delay at that speed is 3.33 m of driving blind.
More Jev resources like this
4GitHub stars
A Craftax (Crafter) survival harness where deterministic code lists every feasible action with its facts and Jev picks one Choice per step, optionally guided by an LLM-written objective and standing rules; three-seed ablations are included.
Jev Demos & Experiments#community#craftax#games
Community
mansicer/jev-playsPython
mansicer/jev-playsPython1GitHub stars
MuJoCo simulation where a local decider-2b model picks the next skill for an AgileX PiPER arm through a Jev-style typed-choice interface while the red cube target is pushed to new positions mid-run. Deterministic checks may reject a chosen skill but never substitute another.
Jev Demos & Experiments#community#robotics#simulation
Community
Hu-xiao-max/jev_robotPython
Hu-xiao-max/jev_robotPython0GitHub stars
Three.js simulation where an OpenAI model turns each plain-language instruction into a plan and Jev makes every reflex decision: two Score questions for move and turn plus one Noul for jump, with code acting only past a 0.33 dead zone and sending no Jev request while no step is active.
Jev Demos & Experiments#community#threejs#simulation
0GitHub stars
Browser Pac-Man race where a deterministic three-junction simulation filters fatal routes, then Jev makes one typed Choice among the surviving route IDs while the server rejects any answer outside the supplied set. Self-hosted Laya runs the same contract for comparison.
Jev Demos & Experiments#community#gaming#simulation
Community
MaryNfs/pacman-ai-raceJavaScript
MaryNfs/pacman-ai-raceJavaScript306GitHub stars
Local macOS computer-use automation on Apple silicon using on-device laya typed decisions and CUA-S1-FORMS for filling, via the Accessibility API.
Jev Demos & Experiments#community#browser#local
Community
FluidInference/FluidUseSwift
FluidInference/FluidUseSwift221GitHub stars
An agent skill that reproduces Jev-style single-step logits over indexed browser controls using local Qwen weights for offline browser use.
Jev Demos & Experiments#community#browser#open-models
Community
APUS-AI-Lab/fast-browser-usePython
APUS-AI-Lab/fast-browser-usePythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.