DEMO / RECORDING
Overview & result
RoboJEV is a two-stage MuJoCo robot controller: Jev selects intent and bounded Cartesian/gripper commands, while code executes and independently checks physical success. The repository publishes both successes and failures.
From the original post
RoboJEV is a small, inspectable robotics laboratory.
How Jev fits in the loop
- Observe robot and object poses, contacts, velocities and recent actions.
- Ask Jev for immediate intent, then XYZ directions and gripper command.
- Execute via Cartesian control and validate physical task conditions independently of model answers.
How to reproduce
- Clone the repository and run the documented MuJoCo/Franka setup.
- Repeat the fixed-seed task evaluation with Jev and the independent rule baseline.
- Inspect physical success checks, failed trials and the original gate-task success/failure videos.
Why this build matters
A useful robotics case with explicit failure evidence and a real rule baseline, not a success-only demo.
Limitations
- Jev receives structured simulator state, not images, and the project has not been validated on physical robot hardware.
- Ten trials per task leave wide uncertainty; the rule baseline outperformed Jev on the gate task.