Jev Robot
MuJoCo PiPER arm controlled by a local decider-2b model choosing among ten skills via a Jev-style typed-choice interface while targets move.
Overview
Jev Robot is an independent simulation project in which a local decision model controls an AgileX PiPER arm while its target moves. It uses the open Mapika/decider-2b model, not TypeSafe's weights, through a Jev-style typed-choice interface. In MuJoCo, a red cube is pushed to three random positions during execution; the robot observes again, recovers from interrupted motions or missed grasps, and places the cube in a tray.
Sensor-derived geometry, visibility, contact and execution history are described in short English facts; the model never sees images or simulator poses. It returns one of ten fixed skills with the full probability distribution from a forward pass, without generating text. Deterministic checks can reject a choice but never substitute another skill, and object motion or unexpected contact triggers a new decision while physics continues. The README reports the frozen controller passed 10 of 10 runs on unused seeds with 30 of 30 relocations, and a 32.6 ms median decision latency on a laptop GPU, noting these are simulation results, not hardware reliability claims.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of ten skills should the arm execute next? | — |
How it works
- Perceive the scene and describe facts in English
- The local decider-2b model picks a skill with probabilities
- Deterministic checks accept or reject the choice
- Execute with IK and joint control; re-decide on disturbance
Strengths
- Fully local, fast decisions (about 33 ms)
- Checks can veto but never substitute choices
- Recovers from moving targets and failed grasps
Caveats
- Simulation only
- Uses a community model, not TypeSafe Jev
- Needs a capable GPU and separate environments
Getting started
Install system rendering libraries, set up the simulation and model environments as described, then follow the Install section.At a glance
- Type
- Agent skill
- Status
- research
- Language
- Python
- Runtime
- MuJoCo simulation with local model
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Robotics control: a local decider-2b model picks the next skill for an AgileX PiPER arm through a Jev-style typed-choice interface while the target moves, with deterministic checks allowed to reject a choice but never to substitute another.
yibie/awesome-jev
Frequently asked questions
Does it use TypeSafe's Jev?
No. It uses the open decider-2b model through a Jev-style interface.
How reliable is it?
The README reports 10 of 10 runs and 30 of 30 relocations in simulation.
How fast are decisions?
About 32.6 ms median on an RTX 5090 laptop GPU.
Related projects
Open 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.
ProjectresearchPython- EmbodiedJev★ 302
Browser MuJoCo Franka Panda workbench comparing Jev, Claude, OpenAI-compatible and local models on pick, stack and obstacle tasks.
ServiceresearchPython 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.
AppprototypeTypeScriptExperimental robot loop where Jev chooses the next verified action from perception-generated options for a language goal.
ProjectresearchPython