The CMU Rover

Hans P. Moravec
Robotics Institute
Carnegie-Mellon University
Pittsburgh, PA 15213

February 1982


@heading(Abstract)

	An advanced camera-equipped mobile robot is being built at the CMU
Robotics Institute to support research in control, perception, planning and
related issues.  It will be used initially to continue and extend visual
navigation work (the Cart Project) completed in 1981 at Stanford's AI Lab,
but is designed with mechanical, sensor and controller flexibility which
should permit it to support a wide variety of other efforts.

@heading(Introduction)

	Experience with the Stanford Cart @cite(Moravec80), a minimal
computer controlled mobile camera platform, suggested to me that, while
maintaining such a complex piece of hardware was a demanding task, the
effort could be worthwhile from the point of view of AI and vision research.
A roving robot is a source of copious and varying visual and other sensory
data which force the development of general techniques if the controlling
programs are to be even minimally successful.  By contrast the (also
important) work on disembodied visual data often focuses on relatively small
image sets, and is in some danger of overspecializing.  Drawing an analogy
with the natural world, I believe it is no mere co-incidence that in all
cases imaging eyes and large brains evolved in animals that first developed
high mobility.

	Other groups have come to similar conclusions @cite(Raphael76)
or have developed sophisticated rovers with an eye to near term applications
 @cite(Carlisle81) @cite(Yakimovsky78).  The new Unimation rover
 @cite(Carlisle81) is very similar in size, shape and mechanical
capabilities to the machine we are building.  It achieves a full three
degrees of freedom of floor-plane mobility by use of three novel
"omnidirectional" wheels which, by virtue of rollers in place of tires, can
freely move in a direction broadside to the wheel plane, as well as
performing, under motor control, the usual wheel motion. We achieve the same
effect with steerable wheel assemblies, described below.

@heading(Overview)

	The project so far has been focused on developing
a smoothly functional and highly capable vehicle which will serve a
wide variety of future research.  A hint of this future work can be found
in @cite(Moravec80) and later in this paper.

	The shape, size, steering arrangements and onboard and external
processing capabilities of the rover system were chosen to maximize the
flexibility of the system (naturally limited by present day techniques).

	The robot is cylindrical, about a meter tall and 30 cm in diameter 
(Figure 1) and has three individually steerable wheel assemblies which give
it a full three degrees of freedom of mobility in the plane (Figures 2, 3).
Initially it will carry a TV camera on a pan/tilt/slide mount, several short
range infrared and long range sonar proximity detectors, and contact
switches. Our design calls for about a dozen onboard processors (about half
of them powerful 16 bit MC68000s) for high speed local decision making, servo
control and communication (Figure 4). Serious processing power, primarily
for vision, is to be provided at the other end of a remote-control link by a
combination of a host computer VAX 11/780 an FPS-100 array processor and a
specially designed high performance analog data acquisition and generation
device.

	We hope eventually to provide a manipulator on the rover's
topside, but there is no active work on this now.  We chose the high
steering flexibility of the current design partly to ease the requirements on
a future arm. The weight and power needed can reduced by using the mobility
of the rover to substitute for the shoulder joint of the arm.  Such a
strategy works best if the rover body is given a full three degrees of
freedom (X, Y and angle) in the plane of the floor. Conventional steering
arrangements as in cars give only two degrees at any instant.

@heading(Details)
	
	Three degrees of freedom of mobility are achieved by mounting the
chassis on three independently steerable wheel assemblies (Figure 2). The
control algorithm for this arrangement at every instant orients the wheels
so that lines through their axles meet at a common point. Properly
orchestrated this design permits unconstrained motion in any (2D)
direction, and simultaneous independent control of the robot's rotation
about its own vertical axis.  An unexpected benefit of this agility is the
availability of a "reducing gear" effect.  By turning about the vertical
axis while moving forward the robot derives a mechanical advantage for its
motors.  For a given motor speed, the faster the rover spins, the slower it
travels forward, and the steeper the slope it can climb. (Visualization of
this effect is left as an exercise for the reader.)

	To permit low friction steering while the robot is stationary, each
assembly has two parallel wheels connected by a differential gear. The drive
shaft of the differential goes straight up into the body of the robot. A
concentric hollow shaft around this one connects to housing of the
differential (Figure 3). Turning the inner shaft causes the wheels to roll
forwards or backwards, turning the outer one steers the assembly, with the
two wheels rolling in a little circle. The assemblies were manufactured for
us by Summit Gear Corp.

	Each shaft is connected to a motor and a 4000 count/revolution
optical shaft encoder (Datametrics K3). The two motors and two encoders are
stacked pancake fashion on the wheel assembly, speared by the shafts. There
are no gears except for the ones in the differential. (Figure 3 shows a
schematic cross section of a complete motor/wheel-assembly structure,
and a partially assembled stack in the flesh).

	The motors are brushless with samarium-cobalt permanent magnet
rotors and three-phase windings (Inland Motors BM-3201). With the high
energy magnet material, this design has better performance when the coils
are properly sequenced than a conventional rotating coil motor. The coils
for each are energized by six power MOSFETs (Motorola MTP1224) mounted in
the motor casing and switched by six opto-isolators (to protect the
controlling computers from switching noise) whose LEDs are connected in
bidirectional pairs in a delta configuration, and lit by three logic signals
connected to the vertices of the delta.

	The motor sequencing signals come directly from onboard
microprocessors, one for each motor. These are CMOS (Motorola MC146805 with
Hitachi HM6116 RAMs) to keep power consumption reasonable. Each processor
compares the output of its shaft encoder to a desired motion (supplied by
yet another processor as a time parameterized function) and energizes or
de-energizes the appropriate motor phase to keep the motor on track with no
wasted power. Since each motor has a potential power dissipation of 250
watts this is more than a nicety.

	The shaft encoder outputs and the torques from all the motors, as
estimated by the motor processors, are monitored by another processor,
 @i(the Simulator), a Motorola MC68000 (with all CMOS support circuitry the
power requirement for our 32K 68000 is under one watt. The new high
performance 74HC series CMOS allows operation at full 10MHz speed.) , which
maintains a dead-reckoned model of the robot's position from instant to
instant. The results of this simulation (which represents the robot's best
position estimate) are compared with the desired position, produced by
another processor (68000), @i(the Controller), in yet another processor
(68000), @i(the Conductor), which orchestrates the individual motor
processors. The conductor adjusts the rates and positions of the individual
motors in an attempt to bring the Simulator in line with requests from
the Controller, in what amounts to a highly non-linear feedback loop.

	Other onboard processors are:

@begin(description)

Communication @\ A 68000 which maintains an error corrected and checked
packet infrared link with a large controlling computer (a VAX 11/780 helped
out by an FPS 100 array processor and a custom high speed digitizer) which
will do the heavy thinking.  Programs run in the Controller are obtained over
this link.

Sonar @\ A 6805 which controls a number of Polaroid sonar ranging devices
around the body of the rover. These will be used to maintain a rough
navigation and bump avoidance model.  All measurements and control
functions of this processor and the following ones are available
(on request over a serial link) to the Controller.

Camera @\ A 6805 which controls the pan, tilt and slide motors of the
onboard TV camera. This compact camera (Edo Western 1631) broadcasts its
image on a small UHF or microwave transmitter (made by 3DBM Inc?). The
signal is received remotely and the video signal captured by a high
bandwidth digitizer system and then read by the remote VAX. There are
tentative plans for a  minimal onboard vision system using a 68000 with
about 256K of extra memory onboard the rover, for small vision tasks when
the rover is out of communication with the base system.

Proximity @\ A 6805 which monitors several short range modulated infrared
proximity detectors which serve as a last line of defense against collision,
and which sense any drop off in the floor, and contact switches.

Utility @\ A 6805 which senses conditions such as battery voltage and motor
temperature, and which controls the power to non-essential but power hungry
systems like the TV camera and transmitter.

@end(description)

	Communication between processors is serial, via Harris CMOS UARTs,
at a maximum speed of 256 kilobaud. The Conductor talks with the motor
processors on a shared serial line and the Controller communicates with the
Sonar, Camera, Proximity, Utility and any other peripheral processors by a
similar method.

	The processors live in a rack on the second storey of the robot
structure (Figure 1), between the motor and battery assembly (first floor)
and the camera plane (penthouse).  Figure 4 shows the rack alone, not fully
populated.

	The rover is powered by six sealed lead-acid batteries (Globe
gel-cell 12230) with a total capacity of 60 amp hours at 24 volts.  The
motors are powered directly from these, the rest of the circuitry
derives its power indirectly from them through switching DC/DC
converters (Kepco RMD-24-A-24 and Semiconductor Circuits U717262).
Each 6805 processor draws about one eighth of a watt, each 68000
board only one watt.

	Physically the robot is a meter tall and 55 centimeters in diameter.
It weighs 90 kilograms. The maximum acceleration is one quarter g, and the
top speed is 10 kilometers/hour. With appropriate onboard programming the
motions should be very smooth.  The great steering flexibility will permit
simulation of other steering systems such as those of cars, tanks and boats
and other robots by changes in programming.

@heading(Progress)

	As of this writing (June 8, 1982) we are awaiting final fabrication
of the remainder the robot's mechanical parts, and have bench prototypes of
all the major electronics.  Debugging is proceeding furiously. Three
instances of each of the two processors are completely operational.  Low
level control software has been debugged on a rover simulation, but remains
to be tested on hardware.  We feel confident that a minimal demonstration of
a running rover can be arranged for the mid August date of this conference.

@heading(Promises)

	The hope expressed in the introduction, that mobility somehow forces
evolution in the direction of true intelligence, may already be being
vindicated.  Our efforts to devise a suitable programming language useful
for writing Controller scripts for interesting robot tasks floundered until
we realized clearly that controlling a rover is an essentially different
task than controlling a robot arm. An arm, in the typical assembly line
setting, executes a sequence of tasks, possibly with branching, that can be
visualized and charted in advance.  Exceptional conditions are an infrequent
nuisance that can be handled by an error exit or a small number of interrupt
activated routines that reset the situation to some standard state and
return control to the mainstream code.  This approach does not work for
programming a rover in typical environments because @i(almost every
condition encountered by a rover is exceptional). Obstacles and
opportunities cannot in general be anticipated in advance, and if we
attempted to deal with them by interrupt coding, our stacks would soon
overflow as exception interrupted exception.

	We are now working on an approach in which the Controller is
governed by a set of production rules, each rule having an activating
condition involving the state of sensors and internal variables and
consequences which affect internal variables and motor functions.
Conceptually the potentially relevant rules are active at all times, and may
take over control from each other as preconditions change.  Conflicts are
mediated by weights (or confidence levels) produced by activated rules.
Sequences of steps are made possible in this arrangement by the introduction
of sequence variables, which trigger rules and are then changed by those
rules to trigger successive ones. To maintain some coherence of behavior
there can also be variables which affect the preconditions of large classes
of rules - for instance a variable that at a low value causes the robot to
ignore minor signals from its proximity sensors while responding to them
vigorously if high.  A sequence of proximity signals would gradually raise
the value of this variable, and the robot's behavior would change from
boldly pursuing its main goal (whatever that may be) to cautiously feeling
its way around.  We are tempted to give such variables names like "fear" and
"enthusiasm", but will refrain from doing so for the moment.

	Our system must operate in real time, a need which distinguishes it
from typical rule based expert systems.  We currently imagine a scheme in
which, for large jobs, the entire rule system is maintained in the remote
VAX, with efficiently indexed rule subsets likely to be relevant being
dynamically downloaded as conditions change.  Even so it is clear that the
load on the Controller will be heavy. We are preparing a 256 Kbyte memory
expansion board for it, and are looking forward to a challenging job.

	Other interesting preliminary thinking has resulted in a scheme
by which a very simple arm with only three actuators will enable the robot,
making heavy use of its great steering flexibility, to enter and leave
through a closed standard office door.  Discussions with the distributed
sensor net crew here has raised the possibility of equipping the robot with
ears, so it could passively localize sound, and thus perhaps respond
correctly, both semantically and geometrically, to a spoken command like
"Come here!".  We are also toying with the idea of a phased array sonar with
about 100 transducers operating at 50 KHz which, in conjunction with the
high speed analog conversion device mentioned above and the array processor,
would be able to produce a modest resolution depth map (and additional
information) of a full hemisphere in about one second, by sending out a
single spherical pulse, then digitally combining the returned echoes from
the individual transducers with different delay patterns to synthesize
narrow receiving beams.

@heading(Support)

	This work is being lavishly funded by the Office of Naval
Research.