szang18/s3-language-following-v7big

S3 Language-Following v7big

Dual-arm tabletop pick corpus for the S3 language-following benchmark.

• 5,531 episodes over 2,520 scenes (textured rendering, 4 cameras: head / front / wrist×2, 256×256 @10fps, 98 frames/episode)
• Task family: "pick up the apple 〈relation〉 the 〈landmark〉" — 4 spatial relations × 6 landmark objects (18 populated cells), 3 identical apples per scene, 1 distractor landmark, 8 caption phrasings per cell
• Balanced ~62 demos per (cell × distractor-config); scenes with failed/not-lifted demos filtered
• LeRobot v2.1 layout; meta/s3_episode_map.json carries per-episode scene/cell/caption metadata
• State/action: 54-D dual-arm joints; relative-action stats in meta/relative_stats.json

Normalization caveat: GR00T-style relative-action normalization uses raw min/max from relative_stats.json, which scales with dataset size. When training on subsets of this corpus, transplant this corpus's meta/relative_stats.json (and stats.json) into the subset instead of regenerating them, or small-N noise dims will be amplified into an unlearnable loss floor.