Action-Conditioned World Models for Autonomous Navigation
Master's Capstone Research (University of Adelaide) — Supervised by Prof. Peng Shi & Dr. Bing Yan.
Designed and trained a 4.57M-parameter Recurrent State-Space Model (RSSM) and a hybrid Cross-Entropy Method (CEM) planner for mapless robot navigation. Achieved 0.815 SSIM for multi-step depth prediction, reduced collisions by an order of magnitude compared to baseline PPO reactive models, and transferred the model to physical Unitree Go2 quadruped hardware with a proximity-gating safety layer.