Devonance/rover-claude-jev-demo
User repository on GitHub
Just a weekend project with Claude as system two, and Jev as system One.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Rover Claude Jev Demo
Simulated Perseverance crossing real USGS Jezero terrain where Jev answers narrow onboard questions in about 250 ms, Claude writes the sol plan in 15 to 80 s, and code owns every threshold.
- Stars
- 3
- Language
- JavaScript
- Published by
- Community
- Author
- Devonance
- Last verified
- Oct 2026
About this project
This is a demo. It is not flight software, it is not a research result, and nothing here has been validated against a real rover. It is a toy built to make one idea visible: that a robot's decisions split into two kinds, and that you can use two very different models for the two kinds and watch the seam between them.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Jev is text-only by design: depth frames are unprojected to world points, features become numbers, then numbers become words.
- The gate is a single rule: below 0.55 confidence nothing acts, so the rover stops, raises the mast and takes a second stereo pair.
- Code plans nine candidate steering arcs per cycle; every camera pair images at once and any one pair can veto the move.
- Telemetry arrives about once a second and Jev triages battery, thermal, currents, slip, odometry, tilt, dust and their trends.
Watch out
The README opens by stating this is a demo, not flight software and not a research result, and that nothing has been validated against a real rover. The limitations section is measured against the world's rock list, which the rover never sees. It needs Node 20 or newer, Chrome, a TypeSafe key and the Claude Code CLI logged in, and the author quotes roughly $1.10 in Claude usage against $0.03 in Jev for a full run. The text-only path is the real trade: depth is unprojected, features are measured as a stereo pipeline would, and 0.42 m becomes knee-height, so a lot of visual detail is deliberately discarded before the model sees anything. Claude built the decision graph and then got it back as an MCP tool, so the same model that plans also authors the questions it asks. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
No, and the README says so in the first paragraph. It is a simulation of Perseverance driving real USGS Jezero terrain, with the measured limitations section scored against a rock list the rover never sees. Treat it as an argument about where a fast typed model belongs in a decision loop, not as robotics results.
The project's claim is about shape rather than capability. A question asked a hundred times a minute needs a number you can threshold, and the README argues an LLM returns prose whose reported confidence is not calibrated at three orders of magnitude too high a latency. A System One model cannot plan, argue or write, which is the point: it gets structured state and typed questions and returns a distribution in one fast call.
No. Each camera renders a colour frame and a depth frame standing in for a stereo pair, depth is unprojected into world points, relief above local ground is measured, connected blobs become features with distance, bearing, footprint, max relief, tone and shadow fraction, and those numbers are bucketed into words before anything is sent. Stereo fails in deep shadow and in patches rather than pixel by pixel, so the model inherits those failure modes by construction.
The author puts a full run at about 20 minutes and roughly $1.10 of Claude usage against about $0.03 of Jev. The split explains it: Claude is called a few times per sol for planning, interpretation and diagnosis at 15 to 80 s each, while Jev is called hundreds of times per sol at about 250 ms for telemetry triage, obstacle vetoes and target scoring.
Node 20+, Chrome, a TypeSafe key, and the Claude Code CLI logged in; clone the repo, run npm install, put the key in .env, run claude once to log in, then npm start and open http://localhost:4180.
npm run smoke loads the page headless and reports console errors, while npm run autorun runs the full headless sequence and writes screenshots to shots/.
Yes. The same decision graph runs as ROS 2 Jazzy nodes under WSL 2 with the browser as the simulator over rosbridge; ros2/setup_wsl.sh provisions it and ros2/start_graph.sh launches it.
More Jev resources like this
30GitHub stars
Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.
Jev Demos & Experiments#community#jevbest#catalog-import
19GitHub stars
Live versus Tetris where Jev faces Claude Haiku 4.5, Gemini 3.8 Flash, or local open-weight Laya on the same seeded piece sequence; cleared lines arrive as garbage rows and the first to top out loses.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
trungdq88/jev-tetrisJavaScript
trungdq88/jev-tetrisJavaScript21GitHub stars
Browser top-down driving sim that sends sensor JSON to Jev every 200ms with four typed driving questions in one call, then applies confidence gating and reflex rules to steer.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
vinilana/live-jevJavaScript
vinilana/live-jevJavaScript12GitHub stars
Chrome extension that labels each X post with Substance, Humor, Chit-chat, Promo or Junk plus AI-written and off-topic flags, collapsing hidden posts to one line.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bohutang/siftJavaScript
bohutang/siftJavaScript11GitHub stars
Chrome extension that labels each X post with intent plus rage-bait, synthetic, and undisclosed-ad signals from Jev, with threshold-only rendering and local caching.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
manifoldor/xtagsJavaScript
manifoldor/xtagsJavaScript8GitHub stars
Chrome extension that judges Taobao and JD listings against double-weekend-off clues and stamps a PASS mark over product images it guesses come from non-weekend-off shops.
Jev Demos & Experiments#community#awesomejev#catalog-import
From 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.