|
Technological advances in artificial intelligence have created an opportunity for effective teaming
between humans and robots. Reliable robot teammates could enable increased situational awareness
and reduce the cognitive burden on their human counterparts. Robots must operate in ways that follow
human expectations for effective teaming, whether operating near their human teammates or at a
distance and out of sight. This ability would allow people to better anticipate robot behavior after
issuing commands. In comparison to traditional human-agnostic and proximal human-aware path
planning, our work addresses a relatively unexplored third area, team-centric motion planning: robots
navigating remotely in an unfamiliar area and in a way that meets a teammate's expectations.
In this paper, we discuss initial work towards encoding human intention to inform autonomous robot
navigation. Our approach leverages the methodology and data collected in an ongoing series of natural
dialogue experiments where naive participants provide navigation instructions to a remote robot
situated in an unfamiliar environment. Participants are tasked with uncovering specific information
about the environment via the remote robot through real-time mapping and snapshots, requiring fine-
grained robot movement that meets the intention of a given command. This sensitivity often leads to
clarification commands to augment the position and orientation of the robot in order to achieve the
desired instructor intention; we seek to reduce or eliminate the need for these clarification commands
for more efficient task completion. Our current efforts use known participant responses to executed
commands to train a reinforcement learning policy for building awareness about unknown
environments. Ultimately, this approach would lead to robot movement that maximizes the amount of
relevant information relayed back to a human instructor while minimizing instructor burden.
|
