Dataset Card for ARES Data

Author: Jacob Phillips

ARES Github Repository

Dataset Description

This dataset contains the data and databases for the ARES project. It is a collection of 5000 rollouts from the Open X-Embodiment dataset which have been ingested into the ARES platform. The data is stored in three databases:

• SQL database: a SQLite table containing structured information like length, success rates, and esimates provided by a VLM (such as description, focus objects, environment, etc.). The table is stored as robot_data.db and also dumped as a parquet file robot_data.parquet.
• Embedding database: a collection of FAISS indexes holding state and action embeddings for each rollout, as well as embeddings for descriptions and task instructions. The database is stored as a tar file embedding_data.tar.gz.
• Annotation database: a MongoDB collection containing annotations for each rollout, including the rollout name, the rollout data, and the annotations. The database is stored as a tar file annotation_mongodump.tar.gz.

Usage

The data can be downloaded and restored using the release scripts in the ARES repository. First, follow the instructions in the README to set up the environment. Then, run the following command to download and restore the databases:

cd scripts/release
./pull_from_hub.sh