LIBERO-X · LEVEL 1 · Kitchen extension

Successful rollouts, frame by frame

Every task with at least one successful episode (125 of 259), sampled at 8 frames from initial state to terminal state. Scrub a strip, or turn on tracking to follow the arm through the scene.

These frames come from logs/liberox/LEVEL1. results/paper/liberox.json records that sweep as using a different checkpoint and task set, not merged into the paper's LIBERO-X table, so they show this run, not the reported numbers. Tracking marks the largest region of each frame that differs from the static scene; the camera never moves, so the median of an episode is the empty kitchen. It is computed from pixels and is not an object detection or the policy's own grounding, which these logs never recorded. DSL actions, program attempts and env steps are read verbatim from episodes.jsonl.