Small decisions. Interesting possibilities.Submit contentSubmit
embodied-jev preview

embodied-jev

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

Added to Jevfast

How it uses Jev

In the documented layered XYZ setup, Jev receives simulator coordinates and contact feedback, selects a subgoal, then chooses axis directions and gripper action. The controller executes the step and returns updated observations for another decision.

What Jev decides

Atomic motion requestExample answers · not a recorded Jev response · Source ↗
Question 1 · gripper
YOUR APP
INSTRUCTION

Choose the finger command for the current subgoal.

STATE

Measured TCP, object, and destination coordinates plus the selected subgoal, target, signed errors, tolerance, and step limit are sent as state.

JEV · CHOICE
  1. open
  2. hold
  3. close

App workflow

  1. Observe the simulation

    Read structured robot and scene measurements, optionally derived from local visual detection.

  2. Choose a bounded action

    Jev selects a subgoal and then motion and gripper choices from the supplied state.

  3. Execute and observe again

    Apply the action in MuJoCo and feed the resulting state back into the next decision step.

Why it is interesting

The browser-based lab exposes physical feedback and decision stages while supporting configurable cameras and multiple model backends; Jev itself receives structured text rather than camera images.