Hierarchical Approach to Transfer of Control in Semi-Autonomous Systems (Extended Abstract)

Semi-Autonomous Systems (SAS) describe a class of systems that are explicitly controlled by both human and agent actors, incorporating the distinct capabilities of each. They are designed for domains that require quick and safe transfer of control between the agent and human, and embed control transfer decisions in the overall high-level plan. We formally define SAS as a hierarchical model and its properties. We discuss how micro-level transfer of control and macro-level path planning can be solved together. Finally, we explore how SAS can be applied to semi-autonomous vehicles. 1. INTRODUCTION Purely autonomous systems have been applied to a wide range of domains ranging from water reservoir control [3] to energy-conserving smart environments [7]. Almost all applications, however, require human intervention as part of their standard operating procedure. Few models actively incorporate this collaboration, and instead use hard-coded default behaviors [2]. Semi-Autonomous Systems (SAS) are models that explicitly model the collaboration of actors, in order to proactively utilize their respective capabilities [11]. New challenges arise in semi-autonomous systems due to the inherent uncertainty and complexity of human behavior. We consider semi-autonomous driving [10] as our target application. Within this domain, vehicles can only operate autonomous on a subset of the roads (e.g., only well-mapped roads). For longer routes, the vehicle requires the human to occasionally take control during execution. This transfer of control process requires second to second monitoring as the vehicle messages the driver to change the vehicle's controlling entity. It may be unsuccessful due to the state of the human (e.g., distracted) or simply time limitations. The proposed collaborative multiagent framework is quite distinct from other models such as Shared Plans [4], Teamwork [9], and Dec-POMDPs [1]. First, SAS requires exactly one actor in control at a time. Second, transfer of control must be explicitly modeled. Finally, SAS must proactively leverage each actor's capabilities (or lack thereof) as it efficiently moves through the sate space.