ViT-Enhanced Critic in Terrain-Adaptive Bipedal Robots

Created using ChatSlide

This research explores enhancing terrain-adaptive capabilities of bipedal robots through Vision Transformers (ViT) in Reinforcement Learning (RL) critic networks. Motivated by the challenges of underactuation and terrain variability in human environments, the study critiques existing methods that overlook spatial terrain correlations. It proposes a novel ViT-enhanced critic architecture, utilising self-attention for improved spatial modelling. Techniques for reward design and training...