Domino_long

# DOMINO long-memory training data This directory contains the ten datasets from `data/domino_long`, converted to the LeRobot v2.1 layout consumed by DOMINO's PUMA trainer. - Datasets: 10 - Episodes: 500 - Training frames: 221,215 - Camera videos: 1,500 (head, left wrist, right wrist) - FPS: 15 - State/action dimension: 14 - Mixture name: `domino_long_memory_10tasks` Each task directory contains `data/`, `videos/`, and `meta/`. PUMA-specific metadata includes `meta/modality.json` and `meta/stats_gr00t.json`. The source episode's `seen` instruction is selected deterministically and stored through LeRobot's `task_index`/`tasks.jsonl` representation. From `policy/PUMA`, launch with the normal DOMINO PUMA environment and override the dataset root and mixture: ```bash DATA_ROOT_DIR=/mnt/sda/zijian/DOMINO/data/domino_long_lerobot \ DATA_MIX=domino_long_memory_10tasks \ bash scripts/run_scripts/run_lerobot_robotwin_puma.sh ``` Revalidate all converted data: ```bash /mnt/sda/zijian/.envs/causal_pi05/bin/python \ /mnt/sda/zijian/DOMINO/script/convert_domino_long_to_lerobot.py --validate-only ``` See `conversion_summary.json` for per-task counts.