Evaluation Assets (GE30 + v6cond)
This folder is a minimal Hugging Face upload payload for updating
UCBProject/eval_assets.
It mirrors repository-relative paths, so after downloading you can sync it into the repo root and the files land in the correct locations.
What this update contains
evaluation/configs/eval_objects_merged_success_ge30.csv- 87 objects (criterion: round≥3
robot_gtcumulativesuccess_count ≥ 30)
- 87 objects (criterion: round≥3
evaluation/configs/README.md- Notes about GE30 being current; GE20/GE40 deprecated
output/pdm/checkpoints_yaw_v6cond/best_model.pthoutput/pdm/checkpoints_yaw_v6cond/pose_stats.ptoutput/affordance_no_rot_executed/min20/checkpoints_v6/best_v6_model.pth
What is intentionally NOT included
This upload does not change heavy assets (mesh/USD/scene files), and does not include evaluation outputs:
data_hub/meshes/SAM3DMesh/rotated_mesh/data_hub/ProcessedData/obj_meshes/*/scale.jsonoutput/obj_usd/sim/assets_scene/,sim/object_rotation_overrides.jsonoutput/evaluation/**(logs/videos/results)
Those stay as-is in the existing dataset version.
How to use after download
Download to a temp directory and sync into the repo root:
huggingface-cli download UCBProject/eval_assets \
--repo-type dataset \
--local-dir /tmp/a2g_eval_assets
cd /path/to/Affordance2Grasp
rsync -a /tmp/a2g_eval_assets/ ./
Notes
- The evaluation object list is now GE30 (
eval_objects_merged_success_ge30.csv). - PDM default checkpoint used by current eval scripts is yaw-conditioned v6cond
under
output/pdm/checkpoints_yaw_v6cond/.