A Probabilistic Inference Approach for Skill-Based Shared Autonomy in Assistive Robotic Manipulation

We present a skill-based shared autonomy approach that addresses the policy blending problem by adaptively arbitrating control between human input and autonomous assistance. Our method uses Bayesian inference to continuously assess user skill from their control inputs and task performance, enabling dynamic adjustment of control authority to match user proficiency. This adaptive approach ensures that experienced users retain control while novices receive appropriate support. We evaluate our method in two teleoperation scenarios involving pick-and-place and cup-stacking tasks with a robotic arm. Simulation and real-robot experiment results show that our method improves task efficiency and success rates while enhancing user satisfaction compared to rigid arbitration strategies.