Hyperparameter sensitivity in imitation learning for manipulation

Characterize and mitigate the sensitivity of imitation learning performance to training hyperparameters across Behavioral Cloning (BC), BC-RNN, and HYDRA on the studied robot manipulation tasks, and develop principled approaches for robust hyperparameter selection.

Background

The paper reports extensive parameter sweeps for multiple methods and environments, observing that performance can vary significantly with choices such as horizon length, network widths, and other training hyperparameters.

The authors explicitly identify the resulting hyperparameter sensitivity as an open problem, indicating a need for systematic methods to understand and reduce this sensitivity across imitation learning approaches in manipulation.