Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design
Summary
Transformer Transformer introduces a diffusion-based model that co-designs robot embodiments conditioned on motion. It unifies embodiment and dynamics through RoboTokens and Dynamics Self-Guidance, enabling zero-shot reward optimization and significantly faster, higher-quality designs compared to traditional optimization baselines, with real-world cloth-flinging experiments validating the approach.