Behaviors for Robust Door Opening and Doorway Traversal with a Force-Sensing Mobile Manipulator

Abstract

Fully autonomous robots will often need to open doors and traverse doorways in order to freely operate within human environments, and assistive robots that open doors on command would potentially benefit the motor impaired. In spite of these opportunities, autonomous manipulation of doors remains a challenging problem after more than a decade of research. Until recently, published research has focused on one or two aspects of door opening, and included results from only a small number of tests on a single door. Within this paper we present a set of behaviors that enable a mobile manipulator to reliably open a variety of doors and traverse doorways using force-sensing fingers and a laser range finder. With this system, a user only needs to briefly illuminate a door handle using a green laser pointer, after which the robot autonomously locates the door handle, finds the manipulable end of the door handle, twists the door handle, and pushes the door open while traversing the doorway. The behaviors use sensory feedback to continuously monitor task-relevant aspects of the world and respond to common forms of variation in the task, such as whether the door is locked or unlocked, is blocked or unblocked, opens to the right or left, or has a handle that twists down clockwise or counterclockwise. We tested the robot in 30 trials with 6 different doors from an initial position over 1.6 meters away from the door handle. For the 24 trials with unlocked doors, the robot succeeded at the entire task in 21 trials (87.5% success rate). In the 6 trials with locked doors, the robot successfully detected that the door was locked in all 6 trials (100.0% success rate). For all 30 trials, the robot stopped in a safe manner without requiring human intervention after detecting failure or success at the task. We conclude with a discussion of how this work relates to several broader issues for intelligent manipulation within human environments, including the use of 3D locations to select behaviors, the generality of serialized sub-tasks, task-relevant features, active perception, force sensing, and methods for scaling systems to handle more tasks of greater complexity.