training/README.md

MAP-Bench training split

35,454 trajectories · 7,184,098 RGB frames · 15 Hz

This split supplies forward RGB videos, base navigation commands, target-object boxes, image-plane Manipulation Anchor Pose annotations, and task instructions. It is the training data for the UniWAM-Base navigation preset. The 15,000 closed-loop evaluation episodes are provided separately under evaluation/.

From the UniWAM repository:

python scripts/download_assets.py --training

The helper downloads the indices, normalization, and text features and verifies and extracts eight archives to MAP-Bench/training/episodes/. All paths in the indices are relative to the index directory. Each episode contains rgb.mp4 and frame_annotations.jsonl; navigation commands are stored either in actions_vx_vyaw.npy or the annotations, as indicated by the index.

Use MAP-Bench/training as UNIWAM_DATA_ROOT in the training guide. The HM3D-derived data are subject to the benchmark license notes.