Article Per Year (5 Year)
p-Index From 2021 - 2026
0.23
P-Index
This Author published in this journals
All Journal Insyst : Journal of Intelligent System and Computation
Marvin Adinata
Universitas Ma Chung
Computer Science & IT

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

 1 
Implementation of Reinforcement Learning Agent Plugins Using A2C and PPO Methods in the Godot Engine Paulus Lucky Tirma Irawan; Marvin Adinata; Mochamad Subianto
Intelligent System and Computation Vol 8 No 1 (2026): INSYST: Journal of Intelligent System and Computation
Publisher : Institut Sains dan Teknologi Terpadu Surabaya (d/h Sekolah Tinggi Teknik Surabaya)

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.52985/insyst.v8i1.470

Abstract

Reinforcement learning (RL) has become one of the most dynamic areas of artificial intelligence, enabling agents to learn optimal behaviour through trial-and-error interactions with their environment. Most RL studies and implementations are conducted using Python-based libraries or within the Unity engine. However, the open-source Godot Engine, which has gained significant traction among indie developers, offers an attractive alternative platform for RL research and interactive game development. This study presents the implementation of RL agents using the Advantage Actor-Critic (A2C) and Proximal Policy Optimization (PPO) algorithms on the Godot Engine through the AgentRL plugin. Two classic RL environments, Cart Pole and Cliff Walking, were replicated in Godot to evaluate algorithm performance. The AgentRL plugin facilitates real-time communication between Godot and Python, allowing model training through the Stable-Baselines3 library. Performance was measured using average episode length and average reward, and statistical analysis was conducted using Welch’s t-test. Experimental results indicate that PPO outperformed A2C in both environments: in Cart Pole, PPO achieved an average of 115.93 steps longer balance duration, while in Cliff Walking, it produced 6.19 steps shorter episodes and 27.93 points higher rewards. These findings confirm that PPO offers better training stability and efficiency than A2C. Furthermore, the results demonstrate that Godot, integrated with AgentRL, is a viable and flexible platform for reinforcement learning research and can serve as a foundation for future studies in AI-driven game development.
Co-Authors Mochamad Subianto Paulus Lucky Tirma Irawan