How it uses Jev
RoboJEV sends structured MuJoCo state to Jev to choose a task intent, then asks for X/Y/Z directions and a gripper command. A Cartesian controller executes the selected motion, while separate physical checks determine task success.
What Jev decides
Gripper command requestExample answers · not a recorded Jev response · Source ↗
Question 1 · gripper
YOUR APP
INSTRUCTION
Which gripper command is appropriate now?
STATE
Structured MuJoCo scene state includes the robot pose, cube, target, held-object state, and measured directional relationships. The source asks X/Y/Z directions and a gripper command together.
JEV · CHOICE
- open
- hold
- close
App workflow
- Measure the scene
Collect TCP and object poses, velocities, jaw width, contact measurements, geometric relationships, and recent actions.
- Choose intent and action
Ask Jev for a task-specific intent, then for axis directions and a gripper command.
- Execute and check
Run the bounded Cartesian step through the controller and MuJoCo, then evaluate task success independently.
Why it is interesting
The two-stage controller separates task intent from bounded motion choices and grounds each step in measured contact and pose state. Model responses cannot declare success.