AIBRIEF
A learned AI model (a diffusion transformer) called Transformer Transformer takes a single motion demo and designs the body (links, joints, motors, controller) to match, cutting the gap between the robot's motion and the demo by 73% on a
A diffusion transformer called Transformer Transformer takes a single motion demonstration and designs a complete robot to match, including every link, joint, motor, and inertial property, then controls what it just built. On a cloth-flinging task, the team's generated embodiment cut tracking error by 73% and max joint speed by 30% against the hand-built ALOHA2 bimanual platform the result was measured against.
The work reframes robot design as a learned problem rather than a hand-engineered one. Embodiment is treated as a co-designed variable: the model generates candidate bodies from a motion goal, then steers that generation using a learned dynamics signal. The same network plays three roles at once, acting as the body generator, the critic, and the controller across different embodiments, trained on a unified tokenization the team calls RoboTokens.
The result was tested across three design spaces: wheeled bimanual arms, quadrupeds, and humanoids, with the team reporting zero-shot optimization of unseen rewards and trajectories and faster runtime than an evolutionary baseline. The paper and a community thread on Hacker News are circulating the work as a step past learned control toward learned body design.
The 73% is the authors' reported result on a single task. Whether the co-design frame holds across harder, real-world manipulation work remains to be tested.