DEMO / RECORDING
Overview & result
EmbodiedJev is a MuJoCo robot-decision workbench. Jev chooses bounded skills or motion subgoals from structured state; local control code checks safety, executes physics actions and evaluates results independently. The repository includes a real Jev hierarchical-decision recording.
Show the original post (Chinese)
每次动作之后重新观察，再决定下一步。模型负责当前选择；控制器负责动作转换与执行，成功由独立的物理条件判定。
How Jev fits in the loop
- Observe robot state, available skills, contact feedback and candidate directions.
- Ask Jev to choose a phase, skill or short XYZ/gripper action from bounded options.
- Validate and execute the motion in MuJoCo, then independently check the physical result and record the trace.
How to reproduce
- Clone the repository and launch the documented local workbench with Python and Node dependencies.
- Run a MuJoCo task with the TypeSafe Jev provider using structured state or RGB-D-derived coordinates.
- Inspect the saved trace, candidate probabilities and independent physics-based success check; compare against the rule baseline.
Why this build matters
Extends Jev's decision-boundary pattern beyond browser agents while preserving a crucial separation between model choice and physical success.
Limitations
- Jev receives structured text, not raw camera images; RGB-D coordinates come from local detectors.
- The Jev hierarchical demo includes a lost grasp and does not prove precise stable placement; rendered playback speed is not API speed.