The NVIDIA Cosmos3-DROID dataset is hosted on Hugging Face and contains structured robotics data, including state observations, action sequences, and video recordings. The repository is organized into success and failure categories, with metadata detailing episode counts, frame rates, and task descriptions.
A custom data pipeline allows for efficient access to the dataset without requiring a full download. By implementing a byte-range Parquet reader, users can extract specific row groups and columns directly from the repository. This approach enables the analysis of joint positions, gripper activity, and Cartesian motion paths.
The system includes a seek-based video processing pipeline that retrieves only necessary temporal segments from episodes. It supports both PyAV and FFmpeg decoding for AV1 video streams, allowing for frame extraction and resizing for downstream tasks. This method optimizes resource usage during the training set preparation phase.
