Maximizing Demonstration Diversity Under Limited Data Budget

Determine effective strategies to maximize the diversity of scenes in expert demonstration data for robot manipulation imitation learning under a limited data collection budget, in order to improve policy generalization without proportionally increasing the number of collected trajectories.

Background

The generalization of imitation learning policies for robot manipulation depends critically on the diversity of training data, particularly the diversity of scenes present in human demonstrations. However, collecting demonstrations across varied environments is costly and time-consuming, making it impractical to simply scale the number of trajectories to achieve diversity.

This creates a tension between the need for diverse, environment-rich datasets and the realities of limited human data collection budgets. The paper highlights that addressing how to most effectively increase data diversity within such constraints is an unresolved challenge that motivates their multiview pseudo-demonstration approach.