[Paper Review] RoboCupSimData: A RoboCup soccer research dataset

Olivia Michael, Oliver Obst+2 more|arXiv (Cornell University)|Nov 6, 2017

Reinforcement Learning in Robotics11 references3 citations

TL;DR

This paper introduces RoboCupSimData, a large-scale, multi-agent robotics dataset derived from 1,125 simulated RoboCup 2D soccer matches involving top-tier teams from 2016–2017. The dataset provides both ground-truth global state and noisy, local perceptual data from each robot’s perspective, enabling research in multi-robot learning, self-localization, and behavior modeling.

ABSTRACT

RoboCup is an international scientific robot competition in which teams of multiple robots compete against each other. Its different leagues provide many sources of robotics data, that can be used for further analysis and application of machine learning. This paper describes a large dataset from games of some of the top teams (from 2016 and 2017) in RoboCup Soccer Simulation League (2D), where teams of 11 robots (agents) compete against each other. Overall, we used 10 different teams to play each other, resulting in 45 unique pairings. For each pairing, we ran 25 matches (of 10mins), leading to 1125 matches or more than 180 hours of game play. The generated CSV files are 17GB of data (zipped), or 229GB (unzipped). The dataset is unique in the sense that it contains both the ground truth data (global, complete, noise-free information of all objects on the field), as well as the noisy, local and incomplete percepts of each robot. These data are made available as CSV files, as well as in the original soccer simulator formats.

Motivation & Objective

• To create a comprehensive, large-scale dataset capturing both global ground truth and local, noisy robot perceptions in multi-agent robotic environments.
• To support research in multi-robot coordination, self-localization, and behavior modeling by providing real-world-like sensor data with uncertainty and partial observability.
• To enable benchmarking of machine learning techniques such as imitation learning, reinforcement learning, and predictive world modeling in a controlled yet realistic simulation setting.
• To overcome limitations of existing datasets by including both high-fidelity ground truth and realistic, agent-specific perceptual data from 22 robots per match.
• To facilitate reproducible research by releasing the dataset in multiple formats, including CSV, original logfiles, and source code for simulator patching and data conversion.

Proposed method

• Modified the rcssserver simulator (v15.3.0) to log visual and body messages from each of the 22 robots per match at 100ms intervals, with variable frequency based on field of view settings.
• Recorded ground truth state (positions, velocities, actions) every 100ms in binary .rcg format, later converted to CSV using a custom C++ tool (rcg2csv).
• Captured local robot perceptions—distances and angles to landmarks and moving objects—via visual sensor messages, stored in .rcv files and converted to CSV using see2csv.py.
• Generated 45 unique team pairings, each played 25 times (10 minutes per match), resulting in 1,125 total matches and 229GB of uncompressed data.
• Provided additional data including robot commands (in .rcl files), simulation parameters, and source code for simulator patches and data conversion tools.
• Structured data into 1125 game files, each containing 22 local perception files (11 per team) and one global ground-truth file, all available in CSV and original formats.

Experimental results

Research questions

• RQ1 How can multi-robot systems learn effective coordination strategies when trained on data with partial and noisy observations?
• RQ2 To what extent can learned models from local robot perceptions generalize to accurate world-state estimation and self-localization?
• RQ3 Can imitation learning and reinforcement learning benefit from a dataset that includes both ground truth and realistic perceptual constraints?
• RQ4 How do different team strategies and behaviors manifest in perceptual data, and can these be modeled or predicted from sensor inputs alone?
• RQ5 What is the impact of varying sensor frequency and field of view on the quality and utility of perceptual data for robot learning tasks?

Key findings

• The dataset comprises 1,125 full-length 10-minute matches, resulting in over 180 hours of gameplay and 229GB of uncompressed data.
• Each match includes 22 local perceptual streams (one per robot) and one global ground-truth stream, providing a complete multi-agent perspective.
• The dataset contains 17GB of compressed CSV data, with additional raw logfiles and command records available in multiple formats.
• The simulator was successfully patched to log visual and body messages per robot, enabling detailed analysis of perception-action loops.
• The data supports a wide range of machine learning tasks, including imitation learning, self-localization, and predictive modeling of team behavior.
• The availability of both ground truth and local perceptions enables direct evaluation of models that must reason under uncertainty and partial observability.