Transformer Transformer introduces a single diffusion‑based transformer that turns a target end‑effector motion into a fully‑specified robot—links, joints, motors, inertial properties, and a controller—optimized for that task. The model uses Robo Tokens, a unified tokenization that represents both the static embodiment and the dynamic state‑action sequence, enabling a single architecture to learn across wheeled, bimanual, quadruped, humanoid, and hand embodiments.

By conditioning on a demonstration motion and a reward function, the network performs motion‑conditioned robot co‑design, cross‑embodiment control, and embodiment generation without separate pipelines. The same weights are trained jointly on a reward‑agnostic dynamics model; during inference, gradients from reward predictions steer the diffusion toward higher‑performance designs—a technique called Dynamics Self‑Guidance.

In experiments, a robot synthesized for cloth‑flinging on an ALOHA2 bimanual platform achieved a 73% reduction in tracking error and a 30% lower maximum joint speed compared to the baseline, outperforming evolutionary baselines in both performance and runtime. The method demonstrates zero‑shot optimization for unseen rewards and trajectories.

Transformer Transformer shows that a unified tokenized representation and a single diffusion transformer can generate, evaluate, and control diverse robots, paving the way for rapid, task‑specific robot creation.