Uncertainty-Aware Imitation Learning for Adaptive Robotic Manipulation

Researchers propose a context-adaptive policy framework for robotic manipulation that combines task-parameterized, robust, and reactive control. The approach uses Learning from Demonstration (LfD) to acquire policies conditioned on robot state and low-dimensional task parameters, addressing limitations of existing dynamical-system-based methods that overlook environmental modulation. The work builds on policy fusion and uncertainty quantification techniques to generate manipulation strategies that adapt to changing context information.