OliverUrbann/HumanoidRobotSoccer

Fall Prediction Dataset for Humanoid Robots

Dataset Summary

This dataset consists of 37.9 hours of real-world sensor data collected from 20 Nao humanoid robots over the course of one year in various test environments, including RoboCup soccer matches. The dataset includes 18.3 hours of walking data, featuring 2519 falls. It captures a wide range of activities such as omni-directional walking, collisions, standing up, and falls on various surfaces like artificial turf and carpets.

The dataset is primarily designed to support the development and evaluation of fall prediction algorithms for humanoid robots. It includes data from multiple sensors, such as gyroscopes, accelerometers, and force-sensing resistors (FSR), recorded at a high frequency to track robot movements and falls with precision.

Using this dataset, the RePro-TCN model was developed, which outperforms existing fall prediction methods under real-world conditions. This model leverages temporal convolutional networks (TCNs) and incorporates advanced training techniques like progressive forecasting and relaxed loss formulations.

Dataset Structure

• Duration: 37.9 hours total, 18.3 hours of walking
• Falls: 2519 falls during walking scenarios
• Data Types: Gyroscope (roll, pitch), accelerometer (x, y, z), body angle, and force-sensing resistors (FSR) per foot.

Use Cases

• Humanoid robot fall prediction and prevention
• Robot control algorithm benchmarking
• Temporal sequence modeling in robotics

Licensing

This dataset is shared under the apache-2.0 license, allowing use and modification with proper attribution, as long as derivatives are shared alike.

Citation

If you use this dataset in your research, please cite it as follows:

"A Large-Scale Dataset for Humanoid Robotics Enabling a Novel Data-Driven Fall Prediction"

Oliver Urbann, Julian Eßer, Diana Kleingarn, Arne Moos, Dominik Brämer, Piet Brömmel, Nicolas Bach, Christian Jestel, Aaron Larisch, Alice Kirchheim,

2025 IEEE International Conference on Robotics and Automation (ICRA)

How to Use the Dataset

To get started with the Fall Prediction Dataset for Humanoid Robots, follow the steps below:

0. Clone the repository

Please make sure that you have installed git large file support (git-lfs) before cloning this repository.

1. Set Up a Virtual Environment

It's recommended to create a virtual environment to isolate dependencies. You can do this with the following command:

python -m venv .venv

After creating the virtual environment, activate it:

• On Windows:

  .venv\Scripts\activate
  

• On macOS/Linux:

  source .venv/bin/activate
  

2. Install Dependencies

Once the virtual environment is active, install the necessary packages by running:

pip install -r requirements.txt

If you have trouble downloading the requirements, check your internet connection. Alternatively, try increasing the pip timeout or upgrading your pip installation:

# Increase the timeout by 120 seconds
pip install --default-timeout=120 -r requirements.txt
# or upgrade pip
python -m pip install --upgrade pip

3. Run the Example Script

To load and use the plain csv dataset for training a simple LSTM model, run the plain_dataset_usage_example.py script (RAM utilisation exceeds 16 GB):

python plain_dataset_usage_example.py

This script demonstrates how to:

• Load the dataset
• Select the relevant sensor columns
• Split the data into training and test sets
• Train a basic LSTM model to predict falls
• Evaluate the model on the test set

To load and use a already prepared dataset, with reduced RAM utilisation, for training a simple LSTM model, run the lightweight_dataset_usage_example.py script (RAM utilisation less than 2 GB):

python lightweight_dataset_usage_example.py

This script demonstrates how to:

• Convert the csv dataset into a memory mapped file
• Load the memory mapped version of the dataset
• Train a basic LSTM model to predict falls
• Evaluate the model on the test set

The script convert_and_load_dataset.py used by the lightweight example demonstrates how to:

• Select the relevant sensor columns
• Split the data into training and test sets

Make sure to check the scripts and adjust the dataset paths if necessary. For further details, see the comments and docstrings within the scripts.

---

license: apache-2.0

---

Dataset Collaboration Initiative

Introduction

In an effort to enhance research, development, and collaboration in the field of humanoid robotics, we are launching a major initiative aimed at unifying and integrating datasets from various humanoid robots. The aim of this endeavour is to foster innovation, improve accessibility and support the wider robotics community.

Objectives

• Dataset Unification
  We aim to create a comprehensive and standardized collection of humanoid robot datasets, enabling researchers and developers to work with diverse data in a more cohesive manner.

• Tools & Guidelines
  To simplify and optimize the utilization of this dataset repository, we will provide essential tools and guidelines, ensuring a seamless experience for contributors and users alike.

• Open & Inclusive Contribution
  All contributions of datasets that meet the minimum requirements are welcome. We encourage participation from a wide range of robotics projects to foster a richer, more diverse collection of datasets.

Join Us!

If you are working with humanoid robots and have datasets to share, we invite you to contribute to this initiative. Together, we can build a valuable resource that propels advancements in humanoid robotics.

For those interested in making a contribution to this project, the method of contact is via the community tab discussions or if you like by writing to oliver dot urbann at tu-dortmund.de.

Stay tuned for upcoming updates with further details, tools, and submission guidelines available at the HumanoidDataset Repository.