Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

EmbodiedJev

Browser MuJoCo Franka Panda workbench comparing Jev, Claude, OpenAI-compatible and local models on pick, stack and obstacle tasks.

ServiceresearchMIT
View on GitHub ↗Live demo ↗

Overview

EmbodiedJev (行知) brings embodied AI experiments to the browser. A MuJoCo simulation of a Franka Panda arm runs three tasks: carrying a cube to a tray, stacking cubes, and carrying an object over an obstacle. Success is judged by physical state, not by the model. A built-in rule baseline runs without an API key, GPU or weights, and models can be plugged in from the UI: TypeSafe Jev, Claude, OpenAI-compatible APIs and a local MiniCPM5-2B.

Each step observes, decides, converts and safety-checks the action, executes and re-observes. Models choose stages or short steps (XYZ direction, step size, gripper action); Jev receives structured text, with coordinates from a local detector, and APIs with native candidate probabilities show them. Candidate actions can be rehearsed in a simulation copy to catch configured collisions and grasp loss. Two or three models can be compared side by side with the same seed. A live demo site and community reproduction gallery exist; documentation is in Chinese. A list summary notes the published Jev comparison was still pending.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich stage or short-step action (direction, step size, gripper) should the arm take next?—

How it works

  1. Observe structured state (optionally camera images)
  2. Model chooses the next stage or step
  3. Controller converts and safety-checks; optional rehearsal in a sim copy
  4. MuJoCo executes; physics determines success

Strengths

  • Runs without keys via rule baseline
  • Side-by-side model comparisons with replay
  • Physics-verified success

Caveats

  • Simulation only
  • Jev comparison results pending per a summary
  • Chinese-first documentation

Getting started

git clone https://github.com/FBddcz/embodied-jev.git, install, then embodied-jev serve --port 8090.

At a glance

Type
Service
Status
research
Language
Python
Runtime
MuJoCo + local web UI
Has code
yes
Has tests
yes
License
MIT
Stars
302

Use cases and integrations

How curated lists describe it

  • Embodied robotics: MuJoCo decision workbench for a Franka Panda arm that compares TypeSafe Jev System One candidate decisions against reactive baselines and LLMs across pick, place, and obstacle tasks.

    yibie/awesome-jev

  • EmbodiedJev: MuJoCo robot decision workbench with MiniCPM5-2B, Jev and compatible model APIs

    wh000wh000/awesome-jev-live

  • MuJoCo Franka Panda workbench where Jev can choose from bounded simulator actions over structured observations; deterministic code owns physics and safety checks, and the published Jev comparison is still pending.

    cobanov/awesome-jev

  • EmbodiedJev: MuJoCo robot decision workbench with MiniCPM5-2B, Jev and compatible model APIs.

    hellogumbo/awesome-jev

  • MuJoCo robot decision workbench in the browser: three simulation tasks, a local small model or a hosted API, and every observe-decide-act step shown.

    MrJev/awesome-jev

Frequently asked questions

Do I need a robot?

No, it runs in MuJoCo simulation in your browser.

Which models can drive it?

Jev, Claude, OpenAI-compatible APIs and local MiniCPM5-2B, plus a rule baseline.

Who decides success?

Physical simulation checks, not the model.

Related projects

  • DriveJev

    Open 4B System I driving model that scores behaviours like cruise, stop, yield and brake from three camera frames each decision.

    ProjectresearchPython
  • 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.

    Agent skillresearchPython
  • jev-askable-arm

    Simulated Franka arm that follows plain-English goals by having Jev chain about 30 hardcoded motion primitives, one per step.

    Agent skillprototypePython
  • jev-drone

    Camera-only MuJoCo quadrotor where Jev makes 2.5 Hz tactical judgments while code keeps flight control and a safety veto.

    ProjectprototypePython
  • jev-reflex-autonomy-lab

    Multi-drone simulation where Jev makes fast reflex decisions per drone and an optional slower LLM may advise but never takes control.

    AppprototypeTypeScript
  • JEV Control Your Roboarm

    Experimental robot loop where Jev chooses the next verified action from perception-generated options for a language goal.

    ProjectresearchPython