The idea
Dmytro Hrybov tests Jev in a MuJoCo robotics simulation. After the first approach struggled, the experiment split each update into two calls: one for what to do next, and another for the arm and gripper movement. Jev receives simplified geometry and contact information as text.
Inside the decision loop
Text geometry + contact state
Choose a task, then a movement
Update the simulated arm and gripper
What to take from it
A compact state description can matter as much as the choice of model. Separate high-level intent from low-level control and keep the simulator responsible for applying valid movements.
This demonstration uses a simulation and structured text, not direct image input. It does not establish reliability or safety for a physical robot.
Follow the original work
This case is an editorial interpretation of a public community demonstration. It has not been reproduced or benchmarked by this publication.
Original post by @dimentary