Fix off-by-one in the previous re-render commit: the replay video starts with the reset frame, but dataset frame k shows the state after action k, so the videos were one frame behind the states/actions. Now uses render frames [1, 1+length): frame-difference correlation with the original videos peaks at shift 0 for 100/100 episodes (was +1). Only videos change; codec, resolution, fps, frame counts and timestamps unchanged.
Re-render g1_wholebody_bend_pick_mp videos (episodes 243-342) with per-episode table materials
Same videos as SIMPLE-org/data commit a4c4f1c (data-fixed/G1WholebodyBendPickMP-v0.zip): egocentric videos re-rendered so the table/ground materials match each episode's environment_config. Only the 100 mp4 files change; parquet, meta and stats (no image stats) are untouched; codec, resolution, fps, frame counts and timestamps are unchanged.
MP tasks: waist state follows the AMO convention (last commanded value); README: teleop vs MP conventions
