LAB / 023
Embodied AI practice (12): what should the gripper do after recovery is exhausted?
12 MuJoCo rollouts and synchronized three-view replay compare hold, open and open-retreat, separating terminal decisions from physical consequences.
EXPERIMENT RESULT
01 / QUESTION
12 MuJoCo rollouts and synchronized three-view replay compare hold, open and open-retreat, separating terminal decisions from physical consequences.
METHOD
Four observation conditions × three exits, 12 seconds per rollout and 2ms physics steps, keeping E10 gates and budgets unchanged.
RESULT
All three exits fail placement under repeated/permanent gaps: hold retains finger contact, both opening policies drop the cube to the floor.
OBSERVATIONS
Repeated gaps abort on budget at 5.422s; permanent gaps on deadline at 5.142s. Twelve pre-exit trajectory comparisons match exactly.
CONCLUSION
Separate the terminal decision from exit execution. No policy recovers placement in interrupted cells; exit is not task completion.
LIMITATIONS
Fixed simulation, not hardware safety evidence. Lowering silence bypasses transfer-only monitoring. No regrasp, ROS2 or completed VL01.
REPRODUCE
Pinned code and raw records are in the repository; the article provides run, audit and replay commands.
KEEP EXPLORING
Related content
- Engineering cases
Embodied AI practice (12): what should the gripper do after recovery is exhausted?
12 MuJoCo rollouts and synchronized three-view replay compare hold, open and open-retreat, separating terminal decisions from physical consequences.
- PROJECT / 003
Hohoo Embodied Experiments
Two MuJoCo teaching scenes—pick-and-place and planar force control—record contracts, tuning, clearance and delayed-feedback outcomes, including failures.
- Engineering cases
Embodied AI X: why can a recovery budget stop a recoverable task?
Twelve rollouts compare unlimited recovery, a two-resume cap and a 400 ms cooldown, separating bounded behavior from task completion.
- Engineering cases
Embodied AI VIII: what should the gripper do after a drop alarm?
Fifteen fixed-scene rollouts compare hold, open and open-then-retreat after identical alarms.
- Engineering cases
Embodied AI practice (11): placement completed, but is it still valid?
Seven MuJoCo rollouts and a paired replay separate historical completion from current validity, including false alarms and expiry boundaries.