RoboJEV
Two-stage Jev control of a MuJoCo Franka Panda: choose an intent, then X/Y/Z directions and gripper, with physics-based success checks.
Overview
RoboJEV is a small, inspectable robotics lab. Jev receives structured simulator state rather than images and makes two Choice decisions per step: first an immediate intent, then X/Y/Z motion directions and a gripper command. A Cartesian controller executes the action with real MuJoCo contacts, malformed responses are rejected, and each task has independent physical success checks that the model cannot declare for itself.
Tasks are pick and place, surface push, stacking on a fixed pedestal, loose-fit peg insertion and gate obstacle pick and place. Across 100 frozen-seed episodes the README reports Jev at 10/10 on pick and place, push and peg insertion, 8/10 on stacking and 5/10 on the gate task, against an independent rule baseline (10/10 on four tasks, 8/10 on the gate), with Wilson confidence intervals and a reminder that ten trials per task is a small sample. Success and failure videos, an interactive showcase and bilingual docs are provided. Real hardware is a future step.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What immediate intent should the arm pursue now? | — |
| Choice | Which X, Y and Z direction and gripper command should be applied? | — |
How it works
- Read structured simulator state
- Jev chooses an intent
- Jev chooses X/Y/Z directions and a gripper command
- Validate the response and execute via a Cartesian controller
- Check task success independently with physics
Strengths
- Model cannot claim success; physics checks decide
- Failures kept in the denominator and documented
- Rule baseline for honest comparison
Caveats
- Simulation only so far
- Small samples (10 seeds per task)
- Jev underperforms the rule baseline on harder tasks
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- MuJoCo simulation
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Robotics simulation: uses two-stage Jev `Choice` decisions over structured state to select intent and Cartesian motion/gripper commands for a Franka Panda in MuJoCo, rejecting malformed responses and checking task success independently through physics.
yibie/awesome-jevRoboJEV is a small, inspectable robotics laboratory. JEV receives \*\*structured simulator state, not images\*\*, selects an immediate intent, then selects X/Y/Z directions and a gripper command. A Cartesian controller executes the action using real MuJoCo contacts. Each task has independent physical success checks; model answers cannot declare success.
logicrw/awesome-jev-projectsTwo-stage JEV control of a Franka Panda in MuJoCo.
hellogumbo/awesome-jevTwo-stage control of a Franka Panda in MuJoCo from structured simulator state, with physical success checks the model cannot declare for itself.
MrJev/awesome-jev
Frequently asked questions
What does Jev decide?
An intent, then motion directions and a gripper command, from structured state.
How well does it do?
The README reports 10/10 on three tasks, 8/10 on stacking and 5/10 on gate pick and place, with a rule baseline doing as well or better.
Does it see images?
No, only structured simulator state.
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.
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 skillresearchPython