CoolFace
Datasetpublic

ajaysri/shape_hole_v2_visual_trace_preview

Shape Hole V2 Visual Trace Preview This preview shows the low-level-policy variant where the subtask signal is drawn directly onto the external front_workspace_60 image instead of being provided as spatial text. No text is rendered into the image. The training image should match the normal visual observation, with only the trace overlay added. Visual conditioning signal: noisy label_source_uv and label_target_uv Source episode:… See the full description on the dataset page: https://huggingface.co/datasets/ajaysri/shape_hole_v2_visual_trace_preview.

sourceHugging Faceupdated 6mo agoView on Hugging Face
0likes3downloads
Dataset Card

Shape Hole V2 Visual Trace Preview

This preview shows the low-level-policy variant where the subtask signal is drawn directly onto the external front_workspace_60 image instead of being provided as spatial text.

  • No text is rendered into the image.
  • The training image should match the normal visual observation, with only the trace overlay added.
  • Visual conditioning signal: noisy label_source_uv and label_target_uv
  • Source episode: /iris/u/ajaysri/shapes_r_hard/scalingup/outputs/data_v2_slurm/20260326_030005_data_v2_sc_loprio/long_horizon/trimmed_500/seed_0/episodes/episode_000000.npz
  • Rendered video: shape_sort_visual_trace_seed0_episode0.mp4
  • Render summary: shape_sort_visual_trace_seed0_episode0.json