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

Introduces the Transformer Transformer framework, which generates all of a robot's physical properties (links, joints, motors, inertia, etc.) optimized for a specific task, along with its control policy, in a single pass.

The core technical components are as follows:

• RoboTokens: A method for representing a robot's Embodiment and Dynamics as a single unified token sequence. This allows a single model to learn across diverse morphologies such as wheeled robots, quadrupeds, and humanoids.
• Diffusion Transformer: Takes motion Demonstrations as input based on RoboTokens and generates the optimal robot design and controller.
• Dynamics Self-Guidance: A technique that avoids overfitting to the reward function by using a dynamics model designed to achieve a specific reward at inference time, guiding the robot's morphology.

In actual experiments, when designing a robot to perform a cloth flinging task, the approach achieved a 73% reduction in Tracking error and a 30% improvement in maximum joint velocity compared to existing models.