Requirements.md

hf_space_mini_tug Requirements

This directory contains the Streamlit application that runs on Hugging Face Spaces and visualizes each episode.

Files

• app.py – Streamlit entrypoint. Downloads the dataset via snapshot_download, exposes an episode selector, video player, and Plotly charts.
• requirements.txt – Python dependencies (streamlit, plotly, pandas, huggingface_hub).
• README.md – Hub metadata (SDK = Streamlit).

Data Contract

app.py expects the dataset raffaelkultyshev/mini_tug_tape_to_bowl to have:

• meta/info.json describing episodes, data_path, and video_path.
• Parquet files with either:
  1. LeRobot-style observation.state columns (list<float>) or
  2. Long-form frame_idx, joint_name, x_cm, etc. produced by hand_pose_pipeline.py.
• RGB MP4s encoded as H.264 / yuv420p / faststart (videos/chunk-000/rgb/...).

Internal Logic

1. get_dataset_revision() uses HfApi.repo_info so Streamlit caches auto-invalidate when the dataset updates.
2. load_episode() reads the Parquet file and, via build_state_dataframe, pivots the long table back into the [wrist_x_cm, ...] format used by the charts.
   • This function enforces accuracy bounds by reindexing frames and rejecting missing joint rows.
3. Plots are generated with Plotly (build_plot) and share the same axis scaling as the offline PNGs.

Accuracy Guarantees

• Streamlit charts display the exact values computed by hand_pose_pipeline.py. No down-sampling is performed.
• If gaps exist, the app surfaces NaN segments directly, ensuring transparency about measurement uncertainty (<3 mm XYZ, <5° angles as validated in scripts/Requirements.md).

Deployment

• Use huggingface_hub.HfApi.upload_folder(folder_path='hf_space_mini_tug', repo_type='space').
• Before deploying, regenerate the dataset_cache/ by running the app locally or deleting the cache to avoid shipping stale data.***