Transformer Transformer: A Unified Model for Robot Co-Design

Transformer Transformer is a diffusion transformer that takes a manipulation demonstration as input and generates a complete robot embodiment optimized for that motion — every link, joint, motor, and inertial property. It runs on RoboTokens, a unified typed-token representation that encodes robot embodiments, states, and actions across wheeled bimanual systems, quadrupeds, and humanoids within a single architecture.

The same trained weights serve three roles: an unconditional robot generator, an embodiment-aware cross-embodiment controller, and a motion-conditioned designer. At inference time, reward-agnostic dynamics predictions are converted into reward-specific value estimates via a procedure called Dynamics Self-Guidance, which steers embodiment diffusion toward higher-reward designs — replacing separate optimizer, critic, and controller pipelines with one model.

The team fabricated one generated design for cloth flinging on an ALOHA2 bimanual platform, cutting tracking error by 73% and max joint speed by 30% versus the original robot. Across three design spaces, the model performed zero-shot optimization on unseen rewards and trajectories, outperforming an evolutionary baseline in both quality and runtime. For robotics engineers, this signals embodiment design moving from a separate hand-engineered step into something learnable jointly with control policies.