Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Reinforcement learning for obstacle overcoming using a three-dimensional humanoid model

Petroulakis Ioannis

Full record

---

Summary

In the realm of artificial intelligence and robotics, the creation of agents capable of effectively overcoming obstacles is a great challenge. Reinforcement Learning has received substantial attention for its capacity to empower machines to learn and adapt within their surroundings, through interaction with their environment. This has led to groundbreaking advancements in the domain of autonomous agents. This diploma thesis embarks on a journey to harness the potential of Reinforcement Learning, with a specific focus on enabling obstacle overcoming through the utilization of a Three-Dimensional Humanoid Model, commencing from a walking learning example. Building on a comprehensive background, encompassing Unity Game Development, the ML-Agents toolkit, the Anaconda environment for streamlined dependency management, and the fundamental principles of Reinforcement Learning and the Proximal Policy Optimization (PPO) algorithm, the stage is set for a deep dive into the challenges of creating a model able to overcome obstacles. Through a series of experiments, the setup and progress are presented, along with the development of a reward function and the observation space for our agents. Changes in the environment are introduced to assess adaptability and resilience of our model, and PPO hyper-parameters are meticulously tuned for best results. This thesis concludes with promising outcomes, showcasing the creation of a fully functional model, adaptable to diverse environments. Furthermore, it outlines future directions for research and development, aiming to further the quest for intelligent agents capable of undertaking difficult challenges.

Available Files

Services