jev-askable-arm
Simulated Franka arm that follows plain-English goals by having Jev chain about 30 hardcoded motion primitives, one per step.
Overview
Askable arm runs zero-shot robot tasks in the ManiSkill simulator with a Franka Panda arm using the Jev API. You give a goal in plain English, and Jev chains a catalogue of about 30 mutually exclusive hardcoded primitives such as open_gripper, up_1cm, hover_over, descend_onto, press_onto and done. Each step sends privileged simulator state (xyz positions, gripper state, object distances and the recent chain, no images) in one System One request asking which primitive to use and which object it targets; a small PD executor performs that motion, and the loop repeats until done.
There is no training, demonstrations or learned policy; the policy is whatever sequence Jev strings together, and Jev never outputs torques or trajectories. With seed 0 on PickCube and ResetButton tasks, the README shows picking up a red cube in 3 picks, placing it on a target in 7, and pressing two buttons in order in 4, at about one second per pick. It was run on macOS CPU via MoltenVK and is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which primitive should the arm execute next? | open_gripper · close_gripper · up_1cm · hover_over · descend_onto · press_onto · done · ... |
| Choice | Which object does the primitive target? | — |
How it works
- User gives an English goal
- Collect simulator state (no images)
- Jev picks one primitive and target
- PD executor runs the motion
- Repeat until Jev picks done
Strengths
- No training or demonstrations needed
- Control stays in code; Jev never outputs torques
- Readable chain of primitives per task
Caveats
- Uses privileged simulator state, not perception
- Limited to what the primitive set covers
- About one second per step
Getting started
`git clone https://github.com/TarunTomar122/jev-askable-arm.git && cd jev-askable-arm && uv sync`, put JEV_API_KEY in .env, then `source scripts/vulkan_env.sh`.At a glance
- Type
- Agent skill
- Status
- prototype
- Language
- Python
- Runtime
- ManiSkill simulator
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Zero-shot English goals on a simulated Franka. Jev chains hardcoded primitives.
AnotiaWang/awesome-decision-modelsUses Jev to chain predefined skills for English-language goals in a ManiSkill robot-arm simulation.
logicrw/awesome-jev-projectsZero-shot English goals on a sim Franka. Jev chains hardcoded primitives.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Which of about 30 motion primitives to run next and which object it targets.
Does it learn a policy?
No; the policy is the sequence of primitives Jev picks.
Does it use camera images?
No, only privileged simulator state.
Related projects
Camera-only MuJoCo quadrotor where Jev makes 2.5 Hz tactical judgments while code keeps flight control and a safety veto.
ProjectprototypePythonMulti-drone simulation where Jev makes fast reflex decisions per drone and an optional slower LLM may advise but never takes control.
AppprototypeTypeScriptOpen 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.
ProjectresearchPythonExperimental robot loop where Jev chooses the next verified action from perception-generated options for a language goal.
ProjectresearchPythonMuJoCo PiPER arm controlled by a local decider-2b model choosing among ten skills via a Jev-style typed-choice interface while targets move.
Agent skillresearchPythonMuJoCo pilot where Jev, Claude Haiku and reactive rules choose pick-and-place actions; Jev matched the rules exactly.
ProjectresearchPython