colosseum/colosseum-challenge

Colosseum Dataset Card

This dataset contains demonstrations for training and testing Imitation Learning based policies, taken from our simulation benchmark Colosseum, which is based on RLBench. The benchmark consists of 20 tasks from the RLBench suite. We implement variations for each task, like camera pose, which try to test generalization capabilities.

Dataset details

The training set consits of 100 demonstrations of the 20 tasks without any variation factor (the vanilla version of the RLBench tasks). Each demonstration consists of frame data from the following 4 camera views:

• Front camera
• Left shoulder camera
• Right shoulder camera
• Wrist camera

For each camera view we collect the following data:

• RGB
• Depth

Note: each frame is recorded at 128 x 128 resolution.

The test set consits of 25 demonstrations of the 20 tasks, each for the factors of variations that are applicable to that task. Each step collects data from the same 4 camera views, at the same resolution.

Dataset structure

The data is distributed as tar.gz files. After downloading each tar and extracting it into a local folder, you'll get a folder structure like the following (e.g. for the task stack_cups):

Each folder contains a suffix (idx), which indicates which variation factor was applied to the simulation, e.g. idx=0 means no variations, whereas idx=2 means Object Color variation applied to the Manipulated Object. You can find a spreadsheet here with the tasks idx for each of the 20 tasks. You can also find what variations are applicable to that task, as it could be that some variations are not active for some task combination.

The pickle file variation_description.pkl contains the language instructions for that task. Below we go deeper into the folder structure for one of the variations. Notice there is a set of folders per each episode/demonstration, and on each folder there are extra folders for each camera view and type of image. There's also a pickle file low_dim_obs.pkl with the low dimensional observation saved by RLBench. The info stored in this pickle comes from this config file in RLBench.

Downloading the dataset using wget and a download link

1. Go to the HuggingFace repo and select the files option:

2. Select the task you want to get:

3. Get the download link:

4. Use curl or wget to get the tar file:

wget YOUR_DOWNLOAD_LINK

Resources for more information

• Paper: https://arxiv.org/abs/2402.08191
• Benchmark Code: https://github.com/robot-colosseum/robot-colosseum
• Website: https://robot-colosseum.github.io

Citation

If you find our work helpful, please consider citing our paper.

@article{pumacay2024colosseum,
  title     = {THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation}, 
  author    = {Pumacay, Wilbert and Singh, Ishika and Duan, Jiafei and Krishna, Ranjay and Thomason, Jesse and Fox, Dieter},
  booktitle = {arXiv preprint arXiv:2402.08191},
  year      = {2024},
}