Computers Elect. Engng Vol. 17, No. 3, pp. 205-215, 1991 0045-7906/91 $3.00 + 0.00
Printed in Great Britain. All rights reserved Copyright © 1991 PergamonPress plc
REAL-TIME ROBOT ARM COLLISION DETECTION FOR
TELEROBOTICS
CLIFFORD A. SHAFFER
Department of Computer Science, Virginia Polytechnic Institute and State University, Blacksburg,
VA 24061, U.S.A.
(Received 10 December 1990; accepted in final revisedform I1 April 1991)
A~tract--The role of real-time collision detection as part of a safety system in telerobotics applications
is discussed. We survey a number of proposed collision detection systems. These methods differ in
hardware requirements, speed and the information available for use when responding to an immiment
collision. In particular, we compare systems based on simulation of the robots' workspace with systems
that react mainly to local sensor information (i.e. proximity sensors). Both approaches to collision
detection are rapidly improving. Hierarchical data structures allow real-time simulation of the environ-
ment with standard microprocessors. Advances in proximity sensor hardware allow for greater infor-
mation about the current state of the workspace. While at the moment most experimental systems tend
to be entirely either simulation or sensor based, future systems are likely to use local sensor input to update
a sophisticated model of the workspace. We also examine how precision in the representation vs collision
detection time and robot arm stopping time affect the size of the safety buffer required by both approaches.
Finally, the problems of gripping and response to imminent collision are addressed.
1. INTRODUCTION
This paper describes recent developments in real-time collision detection systems. Our primary
interest in these methods is their application to a real-time safety system for multiple robot arms
in both teleoperated and machine controlled environments. The methods vary in a number of ways,
such as complexity of the system, time requirements (real-time vs off-line) and the amount of
special-purpose hardware required. In particular, we recognize a distinction between methods that
rely on hardware sensors to continuously gather local information about the environment (which
we will term the proximity sensor method), as opposed to software simulation of movement within
a predefined environment (which we will term the simulation method). The simulation method
typically maintains a geometric world model describing the working environment.
We discuss certain unsettled issues in the implementation of a robot safety system. We examine
how precision in representation, collision detection time and robot arm stopping time affect the
size of the safety buffer required by the safety system. The safety buffer is that region around the
robot arm that must remain clear of other objects. This safety buffer may change size with arm
speed and direction of motion. We also discuss how differences in the information available to the
two approaches (simulation vs sensor input) affects possible responses when faced with an
imminent collision. Our third issue is how the collision detection module affects the operation of
gripping objects by the robot's end effector.
Conceptually, the proximity sensor approach is quite simple. We mount a set of sensors at
strategic locations on the robot arm. Whenever a sensor detects another object moving within its
sensing range, the sensor can report the position of that object. If the distance falls below the safety
buffer threshold, a collision is imminent. The simplicity of this approach (in its most primitive form)
argues in favor of its adoption for a fail-safe, low-level safety system. Currently the major issue
for the proximity sensor method is finding sensors that: (1) can detect objects regardless of their
material composition or color; and (2) can detect objects at a suitable range [1]. Another issue is
that purely sensor-based approaches (i.e. those that store no model of the workspace) have only
local information about the environment, and thus can only make local decisions in response to
imminent collision.
Simulation-based approaches attempt to model the robot arm workspace, along with all
movement of objects within the workspace. At all times, the positions of all objects that can affect
the arms must be known. In some applications we can assume that all objects other than the arms
and objects gripped by the arms are static. In other applications, we will need to receive information
205