Synergizing Natural Language Towards Enhanced Shared Autonomy.

Shared autonomy can be beneficial in allowing users of assistive robots to refine the robot's behavior and ensure it is adapted to their needs. However, current methodologies mostly focus on using joysticks or physical pushes to modify robots' trajectories, which may not be feasible for people with reduced mobility. In this paper, we present our initial work toward voice-based shared autonomy, implementing a language model which can use sequences of verbal commands to understand the intended correction direction. Our fine-tuned model, capable of running locally on a CPU, shows improved efficiency in complex and realistic sentences compared to recent generative pre-trained transformer (GPT) models.