Search Papers | Poster Sessions | All Posters
Poster C110 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Memory based Generalization for Cognitive Robots
Shweta Singh1, Ritesh Shrivastav2, Siddhesh Dorkulkar2, Vedant Ghatnekar3; 1IIIT, Hyderabad, India, 2COEP Technological University, India, 3MIT WPU, India
Reinforcement learning (RL) holds promise for training agents in complex environments, but generalization remains a key challenge. This study focuses on addressing generalization in maze navigation using Proximal Policy Optimization (PPO) with transformer-based models. We develop a custom maze environment in Unity 3D and train agents using PPO integrated with Transformer XL and Gated Transformer XL architectures. Our experiments assess the agent's ability to generalize policies to unseen maze configurations, demonstrating significant improvements in generalization performance. This research contributes to advancing RL for navigation tasks.
Keywords: Deep reinforcement learning (DRL) Proximal policy optimization (PPO) Trannsformer XL (TrXL) Gated Transformer XL (GTrXL)