The Atlas Benchmark

The Atlas Benchmark offers a collection of scripts and functions for evaluating 2D trajectory predictors.

Atlas allows automated systematic evaluation and comparison of the built-in, external and new prediction methods on several popular datasets (ETH, ATC and THÖR) using probabilistic and geometric accuracy metrics (ADE, FDE, k-ADE and FDE, NLP).


The Atlas Benchmark

Important highlights of Atlas include:

  1. Supported import of new datasets (labeled detection streams),

  2. Support for contextual cues in the environment,

  3. Automated calibration of prediction hyperparameters,

  4. Automated parametrized scenario extraction,

  5. Direct interface to the prediction methods.

Installation and setup

Get all the submodules first:

git submodule update --init --recursive

Please make sure to download the models for S-GAN using the following script:

cd sgan
bash scripts/

Make sure to install atlas in your local virtual environmnet:

python3 -m venv atlas-env

or we recommend to use conda

conda create -n atlas-env python=3.7

Afterwards you can run pip to install the requirements (you may need to install swig via apt)

pip install -r requirements.txt

You can verify the installation by running tests/`.

How to use it

The functionality of Atlas is fully described and illustrated in docs/

For a quick intro, head over to the demo/ folder and check out the tutorial notebooks:



Further details on the motivation and implementation of Atlas can be found in the following paper:

  title={Atlas: a Benchmarking Tool for Human Motion Prediction Algorithms},
  author={Rudenko, Andrey and Huang, Wanting and Palmieri, Luigi and Arras, Kai O and Lilienthal, Achim J},
  booktitle={Robotics: Science and Systems (RSS) Workshop on Social Robot Navigation},


The Atlas Benchmark is developed and maintained by Andrey Rudenko, Luigi Palmieri and Wanting Huang.

In case of questions and comments, feel free to drop us a line at [email protected] and [email protected]


the-atlas-benchmark is open-sourced under the Apache-2.0 license. See the LICENSE file for details.

For a list of other open source components included in the-atlas-benchmark, see the file


View Github