TraceGenDroid / README.md
yoonkyojung's picture
Update README.md
c35bc52 verified
metadata
license: cc-by-4.0
task_categories:
  - robotics
  - depth-estimation
tags:
  - tracegen
  - traceforge
  - droid
  - computer-vision
  - manipulation
  - 3d-trajectories

TraceGen – DROID (Derived Subset)

Overview

This folder contains a derived subset of the DROID dataset processed using the TraceForge pipeline as part of the TraceGen project.

The data is released for academic research purposes.

Evaluation Protocol

This dataset defines the official evaluation protocol for the TraceGen benchmark. Models are evaluated on five environments with the following metrics:

  • Mean Squared Error (MSE)
  • Mean Absolute Error (MAE)
  • Endpoint MSE

The official datasets and models are hosted at: 👉 TraceGen Collection

The official leaderboard is hosted at: 👉 TraceGen Leaderboard

Source Dataset

  • Dataset: DROID
  • Original Authors: DROID Authors
  • Original License: Creative Commons Attribution 4.0 (CC BY 4.0)
  • Official Website: https://droid-dataset.github.io/

What Is Released Here

For each episode <video_name>, this release includes:

  • images/: RGB frames extracted from the original DROID videos
  • depth/: predicted depth maps aligned to the released frames (*.png, *_raw.npz)
  • samples/: per-frame TraceForge / TraceGen representations (*.npz)
  • <video_name>.npz: aggregated episode-level metadata and outputs

RGB frames are redistributed in accordance with the original dataset license.

Modifications

The data has been processed using TraceForge, including:

  • temporal subsampling
  • depth-based 3D reconstruction
  • coordinate normalization
  • trace abstraction

License and Attribution

This derived dataset is released under the CC BY 4.0 license.

Appropriate attribution to the DROID dataset is required, and users must indicate that this dataset is a derived work.

Disclaimer

This release does not substitute for the original DROID dataset. The TraceGen authors assume responsibility only for the derived representations.