JEV Control Your Roboarm
Experimental robot loop where Jev chooses the next verified action from perception-generated options for a language goal.
Overview
JEV Control Your Roboarm explores whether a robot arm can carry out new natural-language manipulation tasks without an autoregressive LLM in the execution loop. Wrist RGB-D perception and memory build a scene snapshot, the program generates the currently valid actions and parameters, Jev chooses what to do next, and deterministic validation and a robot driver execute it before observing again.
Jev never sends motor commands, invents coordinates or sees simulator ground truth; it only picks among actions the program has already verified. The hypothesis is that Jev works as a zero-shot semantic decision layer between perception and existing skills such as pick, place and observe. The control loop and Isaac Sim integration are implemented, but the README says a demo video and a controlled Jev-versus-LLM benchmark have not been published yet.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which valid action or skill should the robot do next? | — |
How it works
- Take a natural-language goal and observe the scene with RGB-D
- Generate the list of valid actions and parameters for the current driver
- Jev chooses the next action
- Validate deterministically, execute the bounded skill, observe again
Strengths
- No LLM call per step
- Jev cannot produce invalid motor commands because options are pre-verified
- Composes new tasks from existing skills without task-specific training
Caveats
- Experimental; no public benchmark or demo video yet
- Runs in Isaac Sim; real hardware results not reported
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- Isaac Sim robot control loop
- Has code
- yes
Use cases and integrations
How curated lists describe it
JEV robot decision loop with an experiment scene.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Which of the program-verified actions and parameters the robot should take next.
Does Jev control the motors?
No. Motion and control stay deterministic; Jev only makes the semantic choice.
Has it been benchmarked against an LLM?
Not yet; a controlled comparison is planned.
Related projects
Simulated Franka arm that follows plain-English goals by having Jev chain about 30 hardcoded motion primitives, one per step.
Agent skillprototypePythonCamera-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.
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