Huy Ha

“Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design”

“The paper introduces RoboToken, a tokenization scheme that converts any articulated robot's embodiment (links, joints, motors), states, and actions into a consistent sequence of continuous-valued vectors.”

“Rather than maintaining separate neural networks for design generation, design evaluation, and robot control, Transformer Transformer consolidates all three into a single diffusion transformer.”

“The optimized design 'reduced tracking error by 73% (13.0 → 3.5 cm) and maximum joint velocity by 30% (2.57 → 1.82 rad/s)' compared to the original ALOHA design”

“Lead author, previously published FlingBot (dynamic cloth manipulation) and Fit2Form (generative gripper design). His trajectory from gripper design → dynamic manipulation → full robot co-design positions him as a leading researcher in generative hardware design for manipulation.”

“The training data pipeline required training '128 RL experts' for the quadruped design space alone, with 'each RL policy takes 16 hours on an NVIDIA A100'”