RoboTwin Embodied Video Generation Dataset for WorldArena2.0

This dataset is designed for embodied video generation and evaluation across two main leaderboards and an interactive arena of WorldArena2.0.

0) Dataset Overview

• Leaderboard (test_dataset): Evaluation set for Leaderboard.Extract the directory from test_dataset.tar.gz

Final evaluation results will be synchronized to Leaderboard and Arena(optional) respectively.

---

1) Folder Structure & Modalities

For any episode key (for example episodeK), the following four files are one-to-one aligned and must be used together:

1. data/.../episodeK.hdf5

   • Main action sequence container.
   • Contains joint action, end-effector pose.

2. instructions/.../episodeK.json

   • Language prompt.
   • Contains one field: instruction.
   • use instruction to generate a video dataset named {model_name}_test from test_dataset.

3. first_frame/.../episodeK.jpg

   • Initial visual condition frame.
   • Use this as the first frame when doing generation.

---

2) Quick Start for Inference

To generate a future video for episodeK:

1. Initial Condition: Use first_frame/.../episodeK.png.
2. Text-driven: Extract instruction from instructions/.../episodeK.jsonand use these as prompt along with the corresponding first_frame to inference.
3. Action-driven: Use data/.../episodeK.hdf5 as action/trajectory along with the corresponding first_frame to inference.

Requirement: Models should take the first_frame and instruction/action as input to generate a video set containing 1000(test) videos corresponding to first_frame, contains 1000(test) videos.

---

3) Technical Specifications

Our data is sampled and processed from the RoboTwin 2.0 dataset. For detailed technical specifications, HDF5 structure, and coordinate systems, please refer to the official documentation: https://robotwin-platform.github.io/

---