Learning Action-Conditioned World Models for Robotic Navigation | Dataset Analysis

AI & Machine Learning Research Project - Dataset Analysis Presentation This video presents my preliminary dataset analysis for the capstone project: “Learning Action-Conditioned World Models for Autonomous Navigation in Robotic Systems” Master of AI and Machine Learning University of Adelaide Supervisors: Prof. Peng Shi and Dr. Bing Yan --- What this presentation covers: * Motivation and research problem in robotic navigation * Limitations of traditional approaches (SLAM, A*, reactive reinforcement learning) * Introduction to world models with planning (RSSM + CEM) * Dataset collection using Gazebo and ROS 2 * Data processing and structure (617,000 transitions) * Key analysis findings (environment difficulty, dataset imbalance, action coverage) * Identified challenges and mitigation strategies * Next steps: model training, planner integration, and ablation studies --- Key highlights: * Large-scale simulated dataset with 617K transitions across five environments * Identification of a critical imbalance in cluttered environments * Proposed mitigation strategies for improving reward prediction * Focus on model-based reinforcement learning with planning for robust navigation --- Presentation slides: https://docs.google.com/presentation/d/1FELXeUTquhHMdgX_lJteQhBYuctgufnw/edit?usp=sharing --- About the project: This research explores how world models allow robots to simulate future states before acting, enabling more efficient and reliable navigation in dynamic environments. --- Feedback is welcome on: * Dataset design and analysis * Model architecture choices (RSSM vs transformer) * Planning strategy and CEM horizon selection --- Keywords: AI, Machine Learning, Robotics, Reinforcement Learning, World Models, RSSM, CEM, Autonomous Navigation, Gazebo, ROS2, University of Adelaide