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

# Summary Researchers propose Transformer Transformer, a unified AI model that generates complete robot designs optimized for specific manipulation tasks based on demonstration motions. The model, trained on a unified tokenization of robot embodiments and actions, can design robots with optimized links, joints, motors, and inertial properties—demonstrated by a cloth-flinging robot that reduced tracking error by 73% compared to the original design. The same architecture can span different robot types and uses, employing a "Dynamics Self-Guidance" procedure to optimize for different reward functions without overfitting.