Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-Design

I introduce Transformer Transformer, a unified model that designs complete robots optimized for specific tasks from simple motion demonstrations. Using RoboTokens and a diffusion architecture, the system generates every link, joint, and motor property while simultaneously validating the design through control. Our approach, Dynamics Self-Guidance, allows for zero-shot optimization of unseen rewards, significantly improving performance over traditional evolutionary baselines.