Find the exact grasp and release,
in any robot clip.

Search the fused DROID + Berkeley corpus, then watch the agent run the perception pipeline — gripper segmentation, aperture signal, change-point detection — at runtime to pin frame-accurate manipulation moments.