Grasping Policy Generalization: The Role of Subject Diversity in Training Data

If your robot only learns from one human, it learns their specific quirks, biases, and kinematic constraints. Our latest experiments prove that injecting subject diversity—different hand sizes, approach angles, and speeds—forces the policy to learn the underlying mechanics of the task rather than memorizing a specific trajectory.

This is why Radioception datasets guarantee multi-subject variation for every task primitive.