Skip to content

data.md

Latest commit

 

History

History
50 lines (42 loc) · 2.63 KB

data.md

File metadata and controls

50 lines (42 loc) · 2.63 KB
Raw

Dataset preparation

Download Amazon-FAR/far-prism-data from Hugging Face. The current release contains 129 successful teacher-rollout trajectories (35 box, 32 bin, 34 barrel, 28 ball), precomputed policy commands, matching contact sidecars, object visual/collision meshes, G1 assets and reference-replay videos. This is a filtered subset of the original 137 trajectories. Raw videos and the teacher-training motion bank remain under review.

Run the data script from the repository root (requires the installed simulation environment, curl and standard Linux utilities). It downloads the single 495 MB archive, checks its SHA256, extracts and verifies every file, then checks the object assets and prepares training shards for the visible GPUs:

bash download_data.sh

The script pins dataset revision 4240492ccc6bde73c3c4e3009ab3de4859bdbdf7. Pass an output directory as the first argument to use another location; rerunning reuses the verified archive and checks the existing dataset and shards. No separate shard-preparation command is needed. With a custom output directory, pass the corresponding paths to train_student.sh (see --help). If an older dataset is already present, choose a new output directory; the script preserves existing data and refuses to mix revisions.

The extracted training inputs are:

data/far-prism-data/
  clips.csv                                # Per-clip index
  data/train-student/data/motion_bank/      # Motion NPZs, commands and object map
  data/train-student/data/contact_sidecars/ # Matching contact intervals and points
  data/train-student/data/robot_assets/     # G1 URDF and depth-rendering meshes
  data/train-student/vis/                   # Reference replays
  data/recon-hoi/                           # Object URDFs, meshes and retargeted motions

The package includes 129 visual meshes simplified to 50,000 faces each and 2,580 convex collision meshes. Collision geometry and training URDF physics match the original assets. Visual geometry differs from the original training assets; depth-observation equivalence has not been validated. Original visual meshes are also available in the Hugging Face dataset.

The script generates data/student_shards_wsN/ (with an environment-count suffix when needed), covering all 129 published clips. If you later select a different GPU count or environment budget, the training launcher prepares a matching set automatically. After it succeeds, follow the student distillation instructions to use them. The dataset and generated shards remain local under Git-ignored data/.