Jev robot control
Recorded xArm7 MuJoCo pick-and-place trial comparing Jev with two language-model controllers under the same Cartesian action interface.
Recorded xArm7 MuJoCo pick-and-place trial comparing Jev with two language-model controllers under the same Cartesian action interface.
Each robot-control cycle asks Jev to choose a pick-and-place intent from measured geometry and contacts, then asks a second Jev request for X/Y/Z movement signs and the gripper command. A deterministic executor applies one bounded motion increment.
Choose the next immediate pick-and-place intent from actual geometry and contact feedback.
Structured MuJoCo geometry, signed target-minus-tool errors, alignment and bilateral contact checks, and current action context; not camera pixels.
Choose X-axis motion for the selected intent.
The same measured state plus selected_intent, per-axis current position and target errors.
Choose Y-axis motion for the selected intent.
The same measured state plus selected_intent, per-axis current position and target errors.
Choose Z-axis motion for the selected intent.
The same measured state plus selected_intent, per-axis current position and target errors.
Choose the gripper command for the selected intent.
The same measured state plus selected_intent, per-axis current position and target errors.
Read structured simulator geometry and contact feedback for the current robot state.
Ask Jev for one intent Choice and then a second request for axis and gripper Choices.
Apply one bounded Cartesian increment, observe the next state, and record responses and trajectories for replay.
The same two-stage Cartesian interface was recorded for Jev and comparison language models; the repository provides replayable answers and physical trajectories instead of only a promotional speed claim.