User repository on GitHub
Zero-shot English goals on a sim Franka. Jev chains hardcoded primitives.
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Cookbooks
Askable arm
A ManiSkill demo where plain-English goals drive a simulated Franka Panda: Jev picks one primitive per step from a catalog of about 30 hardcoded skills, and Python executes the move. No training or policy.
- Stars
- 14
- Language
- Python
- License
- MIT
- Reading time
- 4 min
- Published by
- Community
- Author
- TarunTomar122
- Last verified
- Sep 2026
About this project
Zero-shot robot tasks with the TypeSafe Jev API. Give a goal in plain English. Jev chains a set of hardcoded primitives. The arm does the thing. No training. No demos. No policy. Jev never outputs torques or a trajectory. It picks one primitive per step. Python runs that move. Repeat. On ManiSkill with a Franka Panda (PickCube-v1, ResetButton-v7), seed 0: Type anything that the primitive set can cover, e.g. Move the fingertip upward by at least 4 centimeters. Press the red button first, then press the blue button.
Quoted from the project's README. Read the full README.
Highlights
- On ManiSkill's PickCube-v1 and ResetButton-v7 at seed 0, goals resolve into chains of 3 to 7 Jev picks, about 1 second each.
- One System One request per step chooses the primitive and, when the skill needs a target, the object; Jev never outputs torques or a trajectory.
- The catalog holds about 30 mutually exclusive skills including open_gripper, up_1cm, hover_over, descend_onto, press_onto, and done.
- State sent to Jev is privileged sim data - xyz, gripper open/closed, object distances, and the recent chain - and no images go to the model.
- Click-to-talk uses the browser's on-device speech recognition, so spoken goals are not sent to OpenAI.
Quickstart
bash
git clone https://github.com/TarunTomar122/jev-askable-arm.git
cd jev-askable-arm
uv sync
cp .env.example .env # put your JEV_API_KEY in here
source scripts/vulkan_env.sh
uv run python scripts/ask_arm_server.pyWatch out
MIT-licensed; needs macOS with MoltenVK for SAPIEN rendering, uv, and JEV_API_KEY or TYPESAFE_API_KEY, and this is simulation with privileged state rather than a real robot or a camera model.
Questions about this entry
Short answers drawn from the linked project or page.
Privileged state. What goes to Jev is simulated xyz, gripper open or closed, object distances, and the recent chain; no images are sent. The README says this plainly: perception is privileged state, not a camera model, and this is a simulation rather than a real cell. Jev never outputs torques or a trajectory either, it picks one primitive per step and a small PD executor drives pd_ee_delta_pos.
The README only documents macOS, on CPU through MoltenVK, and that is what the reported runs used. Setup is uv sync plus sourcing scripts/vulkan_env.sh, with MoltenVK installed so SAPIEN can render. No Linux path is described, and there is no statement about GPU-backed rendering, so treat the platform support as untested beyond what is written.
No, and the README says the primitive set is the whole skill library. About 30 mutually exclusive skills cover the space, so you type any goal they can express, such as moving the fingertip up at least four centimeters or pressing red then blue. If the catalog genuinely cannot express something, the fix is adding a primitive, not rephrasing, since Jev only ever chooses from that menu.
Because button episodes are not killed on the first press in this repo, which is exactly why that two-press sequence can complete where an early-terminating variant could not. The reported numbers come from ManiSkill PickCube-v1 and ResetButton-v7 at seed 0, where goals resolve into chains of three to seven Jev picks at roughly a second each. Success rates across seeds are not documented.
MIT-licensed. It is a uv project: `uv sync`, copy .env.example to .env with a JEV_API_KEY or TYPESAFE_API_KEY in it, then `source scripts/vulkan_env.sh` so SAPIEN can render through MoltenVK. For the spoken demo run scripts/ask_arm_server.py and open http://127.0.0.1:8765; the headless suite is scripts/run_jev_ask.py --seed 0, writing results to outputs/jev_robotics/ask_arm/results.json.
Into the browser and nowhere else. Click once and speak, and the page uses the browser's on-device speech recognition with live captions; when you pause, that sentence becomes the next goal, and the loop keeps going with the next one. Nothing is sent to OpenAI. The typed lab UI at http://127.0.0.1:8765/lab skips speech entirely.
A JSON result file under outputs/jev_robotics/ask_arm/ from scripts/run_jev_ask.py --seed 0. The documented figures are chains of three to seven picks at roughly a second each on ManiSkill PickCube-v1 and ResetButton-v7 at seed 0: hover_over, descend_onto, close_gripper is three picks for the red cube, while placing it on the green target takes seven. Success rates across seeds are not reported.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: Pick Verify: a single decision at the warehouse pick station catches confusable picks.
Show HN: Verifying robot picks with a decision modelRead the thread on Hacker NewsX: A dual-arm robot with Jev in the middle layer
YouTube: Dan Makes Things drives an ESP32 robot with Jev decisions.
More Jev resources like this
A MuJoCo and Franka Panda workbench where Jev chooses the next manipulation step for pick-and-place, stacking, and obstacle-carry tasks, with rule baselines and side-by-side model comparison.
Jev Cookbooks#community#python#robotics
Community
FBddcz/embodied-jevPython
FBddcz/embodied-jevPythonA fine-grained robot control study loop for LIBERO tasks: Jev layers an intent, a contact or motion family, and one of 27 control inputs, while reversible MuJoCo previews evaluate each candidate before execution.
Jev Cookbooks#community#python#robotics
Community
Dimweaker/jev-liberoPython
Dimweaker/jev-liberoPythonA camera-only MuJoCo quadrotor that uses Jev for tactical maneuvers at 2.5 Hz while classical perception and a 50 Hz reflex layer keep safety outside Jev, clearing a five-station course.
Jev Cookbooks#community#python#robotics
Community
RomanSlack/jev-dronePython
RomanSlack/jev-dronePythonA side-by-side T-Rex game where local Laya runs on Apple Silicon through MLX and hosted Jev uses the TypeSafe API, showing choices, latency, survival streaks, and crash replays on one course.
Jev Cookbooks#community#python#games
Community
virajbhartiya/laya-vs-jevPython
virajbhartiya/laya-vs-jevPythonTwo-stage Jev control of a Franka Panda in MuJoCo: Jev selects a task intent, then X/Y/Z directions and a gripper command executed against real contacts, with 43 of 50 Jev episodes succeeding versus 48 of 50 for a rule baseline.
Jev Cookbooks#community#python#robotics
Community
lykycy123/RoboJEVPython
lykycy123/RoboJEVPythonTwo WebAssembly builds of the Chocolate Doom 3.1.1 engine play the same Freedoom map from the same initial state, with a Jev Choice decision picking each tactical macro and a local controller turning it into inputs.
Jev Cookbooks#community#typescript#games
Community
lukaske/jev-doom-agentTypeScript
lukaske/jev-doom-agentTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.