README.md

metadata

license: apache-2.0
pretty_name: GTAVBench
language:
  - en
tags:
  - gui-agents
  - game-environments
  - multimodal
  - reinforcement-learning
  - arxiv:2611.13028
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*.parquet

GTAVBench

150 GTA V tasks for evaluating agent perception, reasoning, and control

Code & environments · Paper · Try the benchmark

GTAVBench is the evaluation benchmark of GTAGym, an environment and framework for evaluation and reinforcement learning on GTA V tasks. It combines interactive 3D environments reconstructed from main-story missions with image-based shooting and sniping tasks using GTA V imagery.

Start here: Highlights · Play online · Download · Schema · Task distribution · Citation

🌟 Highlights

• 150 tasks, 11 categories: 103 interactive 3D tasks span navigation and interaction; 47 image-based tasks cover shooting and sniping.
• Ready to explore: each task includes its full English instruction, a 1280 × 720 initial observation, movement labels, and a direct browser-play link.
• Shared evaluation interface: screenshots and keyboard–mouse actions support both direct JSON evaluation and the Codex harness, with automatic outcome checks in the environment.

🎮 Play online

Open all 150 playable tasks →

Try a task	Link
On-foot navigation	C02-N15
Road navigation	C01-N09
Sniping	C11-N01

Every dataset row has its own play_url. Open it, read the instruction, and press Start task. Loading and reading do not consume the task budget. The game renders in your browser; a desktop keyboard and mouse are recommended.

📦 Download and load

Item	Value
Split	test
Tasks	150
Format	Parquet with embedded screenshot bytes
Data file	test-00000-of-00001.parquet
Initial observation	1280 × 720, Hugging Face Image feature
Fields	7, including per-task play_url

pip install datasets

from datasets import load_dataset

tasks = load_dataset("OpenMOSS-Team/GTAVBench", split="test")
task = tasks[0]

print(task["id"], task["category"])
print(task["instruction"])
print(task["play_url"])
image = task["initial_observation"]  # PIL image, 1280 × 720

For authenticated access, log in with hf auth login before loading.

Download the Parquet file with the Hugging Face CLI

hf download OpenMOSS-Team/GTAVBench \
  data/test-00000-of-00001.parquet \
  --repo-type dataset --local-dir ./gtavbench

🧩 Dataset schema

Field	Type	Description
id	string	Task ID in cXX-nYY format: category code and task number.
category	string	Human-readable task category.
instruction	string	Full English instruction used for evaluation.
initial_observation	Image	Initial screenshot before the first agent action.
environment_type	string	3d for interactive environments; image for shooting and sniping.
movement	list[string]	Movement modes involved in the task.
play_url	string	Link that selects this task in the hosted game.

Initial observations are captured at game elapsed time zero. Sniping starts with an unscoped view. Screenshots retain their original bytes and are embedded in Parquet; they may include the in-game HUD and the instruction overlay used by the harness.

Movement vocabulary

Mode	Meaning
walk	On-foot movement, including stairs, ladders, and climbing
drive	Cars, motorcycles, vans, trucks, and ATVs
swim	Swimming
dive	Underwater navigation
boat	Boats and other watercraft
plane	Fixed-wing aircraft
helicopter	Helicopters

A task can involve several modes: ["walk", "drive"] means it starts on foot and requires finding and driving a vehicle. Aircraft tasks starting on foot include walk with plane or helicopter; the submarine task uses ["boat", "dive"]. Image-based shooting and sniping use []. Labels describe the modes involved, rather than a prescribed action sequence.

🗂️ Task distribution

Category	Environment	Tasks
Road Navigation	3D	20
On-foot Navigation	3D	16
Following	3D	6
Interception	3D	13
Evasion	3D	16
Aquatic Navigation	3D	11
Vertical Traversal	3D	5
Aerial Navigation	3D	8
Indoor Navigation	3D	8
Shooting	Image-based	24
Sniping	Image-based	23
Total	103 3D + 47 image-based	150

📊 Benchmark context

The paper evaluates two protocols on the same task catalog:

Protocol	Task instruction	Agent interface
GTAVBench	Textual instruction	JSON keyboard–mouse actions
GTAVBench-AGI	Instruction appears in the first screenshot	Codex harness with the game_step tool

This release provides the test catalog and initial observations. The GTAGym repository contains the executable environments, evaluation runners, task generator, and separate on-foot RL training data. See the paper for benchmark results and RL transfer experiments.

🙏 Acknowledgements

Our primary acknowledgement goes to GTA_SZ and its contributors, whose city environment and game systems underpin GTAGym's 3D tasks. We also credit Rockstar Games for GTA V gameplay imagery and the authors of the supporting software and assets. Full credits and asset-specific notices are in the repository's acknowledgements.

📎 Citation

If you use GTAVBench or GTAGym, please cite the paper:

@misc{chen2026gtagym,
  title         = {GTAGym: Real AAA Game Tasks Environment for Evaluation and Reinforcement Learning},
  author        = {Jiajun Chen and Shicheng Fang and Jiahao Lu and Yuxin Wang and Xipeng Qiu},
  year          = {2026},
  eprint        = {2611.13028},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2611.13028}
}