452 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002
Millibot Trains for Enhanced Mobility
H. Benjamin Brown, Jr., J. Michael Vande Weghe, Curt A. Bererton, and Pradeep K. Khosla, Fellow, IEEE
Abstract—The objective of this work is to enhance the mobility
of small mobile robots by enabling them to link into a train con-
figuration capable of crossing relatively large obstacles. In partic-
ular, we arebuilding on Millibots, semiautonomous, tracked mobile
sensing/communication platforms at the 5-cm scale previously de-
veloped at Carnegie Mellon University. The Millibot Train concept
provides couplers that allow the Millibot modules to engage/disen-
gage under computer control and joint actuators that allow lifting
of one module by another and control of the whole train shape in
two dimensions. A manually configurable train prototype demon-
strated the ability to climb standard stairs and vertical steps nearly
half the train length. A fully functional module with powered joints
has been developed and several have been built and tested. Con-
struction of a set of six modules is well underway and will allow
testing of the complete train in the near future. This paper fo-
cuses on the development, design, and construction of the electro-
mechanical hardware for the Millibot Train.
Index Terms—Distributed robotics, mobility, modularity,
snakes, trains.
I. INTRODUCTION
R
ECENTLY there has been increasing interest in dis-
tributed robotic systems whereby tasks are executed not
by single robots, but by teams of collaborating robots [1]–[4].
Team members may cooperate to explore unknown spaces,
exchange sensor information, provide surveillance data, ma-
nipulate heavy objects, or carry out any of a number of tasks.
Typically small in size, such robots can be maneuverable in
tight areas and well-suited for covert activities. Individual robot
modules may be endowed with specialized sensing, processing,
mobility or manipulation capabilities to complement those
of other team members. The distributed nature of the group
may provide the team with redundant capabilities and/or infor-
mation storage so a single failure does not disable the entire
team. Physically distributed, these robots can provide varied
viewpoints for sensing and perception and broad coverage for
task execution. Potential tasks include surveillance, monitoring,
sample collection, demining, and chemical plume detection.
Manuscript received March 31, 2002; revised October 1, 2002. Recom-
mended by Guest Editors W.-M. Shen and M. Yim. This work was supported
by the Defense Advanced Research Projects Agency’s (DARPA) Distributed
Robotics Program under Grant DABT63-97-1-0003.
H. B. Brown, Jr. and C. A. Bererton are with the Robotics Institute, Carnegie
Mellon University, Pittsburgh, PA 15213 USA (e-mail: hbb@ri.cmu.edu;
curtb@andrew.cmu.edu).
J. M. Vande Weghe was with the Institute for Complex Engineered Systems
(ICES), Carnegie Mellon University, Pittsburgh, PA 15213 USA. He is now
with the Neurobotics Laboratory, Robotics Institute, Carnegie Mellon Univer-
sity, Pittsburgh, PA 15213 USA (e-mail: vandeweg@cmu.edu).
P. K. Khosla is with the Department of Electrical and Computer Engineering
(ECE) Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
pkk@ece.cmu.edu).
Digital Object Identifier 10.1109/TMECH.2002.806226
Fig. 1. Original Millibot, showingtrack base, sonar transmitters/receivers,and
wireless communication.
Development of the Millibots has been underway for a
number of years at Carnegie Mellon, with emphasis on sensors,
communication capabilities and the high-level aspects of
cooperative tasks [5], [6]. As shown in Fig. 1, a small mobile
platform was developed, including: a skid-steered twin-track
base, on-board battery power, an RF communication package,
one or more sensing devices and a microprocessor for low-level
control, sensor data processing and general coordination of
activities. Each Millibot has a top speed of about 20 cm/s on
smooth surfaces, a range of about 30–50 m on a battery charge,
and nominally fits into a 5-cm cube. Typical sensors include:
ultrasonics for obstacle detection and inter-module ranging
up to a meter; IR sensors for short-range obstacle detection
(
10 cm range); a miniature CMOS video camera (0.8- m
sensor) with transmitter; and pyro-electric sensors for detection
of humans and other warm bodies.
The small size of the Millibots provides good maneu-
verability and allows them to operate inconspicuously, but
places severe limitations on their mobility over rough terrain.
Ideally, Millibots should be able to navigate normal outdoor
terrains including grass, dirt, rocky areas, curbs and steps, as
well as indoor environments. The goal of the present work
is to enhance the mobility of the Millibots to allow them to
operate in such areas. The work builds upon those who have
created reconfigurable robot modules [7]–[13], snake-like
robots [14]–[18], specialized stair-climbing robots [19]–[21]
and trains of wheeled or tracked robots [22]–[25]. It has been
shown that step-climbing and ditch-crossing ability are related
to the wheel diameter (or front sprocket diameter for tracked
vehicles), vehicle length and effective friction coefficients with
the contact surfaces [26]. Especially on soft ground, tracked
vehicles provide greater traction than wheeled vehicles [27].
1083-4435/02$17.00 © 2002 IEEE
BROWN et al.: MILLIBOT TRAINS FOR ENHANCED MOBILITY 453
Fig. 2. Possible new modes of locomotion for a train of linked Millibots.
Based on such considerations, we developed the concept of
the Millibot Train, wherein individual Millibot modules with
powered tracks link into a train configuration to negotiate
difficult terrain; then unlink and disperse to perform distributed
activities. This paper focuses on the electro-mechanical design
and construction of the Millibot Train.
II. M
ILLIBOT TRAIN CONCEPT
Fig. 2 shows the basic capabilities envisioned for the Mil-
libot Train. Powered links between the modules allow the train
to conform to the terrain and adopt the desired shape for a par-
ticular task and to lift multiple modules to surmount obstacles.
For example, the train can be locked into a straight shape for
crossing ditches. A rocker shape is useful for traveling on flat or
moderate terrain while facilitatingskid -steeringwith the middle
modules; in this configuration, the reduced contact length mini-
mizes the resistance to yaw motion. For stair climbing, the artic-
ulation actuators lift the first few modules above the nose of the
step. The train can then drive up the steps in a straight configu-
ration, or a traveling wave shape can be adopted to better con-
form to the steps and reduce the traction needed. For climbing
standard stairs, a minimum of seven modules (each approxi-
mately 10-cm long) is required to assure spanning at least two
steps at all times. Cantilever lifting of three modules is needed
to reach the nose of the first step. The concept includes move-
ment in the sagital plane but does not include lateral articula-
tions, since they would add substantially to the size, weight,
and complexity of the modules without greatly improving ob-
stacle-crossing ability.
The original Millibot Train module concept is shown in
Fig. 3. Wide, individually-powered caterpillar tracks provide
drive traction and skid-steering ability. A two-pin coupler and
Fig. 3. Original concept of a Millibot Train module.
latch mechanism with a matching receptacle at the opposite
end allows modules to drive together and dock with each other.
A high-torque actuator drives the coupler articulation for lifting
and holding modules in the desired configuration. The tracks
cover most of the top and bottom and the surfaces between
the tracks are smooth to minimize drag and the possibility
of catching on terrain features. The tracks extend beyond the
top and bottom faces allowing operation upside-down (and in
fact, there is no preferred top or bottom.) Nominal dimensions
of 3 cm high
5 cm wide 10 cm long were selected to
conform approximately to the 5-cm cube specification in terms
of module volume. This flat shape provides attitude stability
and reduces the probability of a module (or the train) becoming
stuck on its side.
This design requires creativity in locating the sensors used on
the original Millibots. Sensors can be mounted to look out from
inside the track loops similar to the ultrasonic sensors shown in
the figure.Openings in thetrack could permit sensors to look up-
ward/downward through the track, or possibly fore/aft through
the sprockets. Sensors might be deployed from the top surface
as needed, then retracted into the body for protection. Small
sensors could be mounted on the coupler—where they could
be tilted by the articulator for increased view—and on the re-
ceptacle at the other end. Another possibility is to have sensor
modules that could be carried on the coupler: for example, the
steering and lift mechanism could provide pan and tilt functions
for a small camera latched onto the coupler. In previous work
[28] we have designed two possible docking systems which may
be integrated into this design to allow for autonomous coupling
of the modules. However, the primary focus of this research is
454 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002
Fig. 4. Train of Millibots climbing a large step. (A) Forward modules being
driven up the step by those still on the ground. (B) Topmost modules lifting and
helping to hold the others against the face to maintain traction. (C) An alternate
position may be assumed if the friction on the top surface is lower. (D) Once the
last module nears the top the forward modules can bend and resume forward
travel.
on mobility, so we will postpone most of the external sensing
and high-level control issues for future work.
III. M
OBILITY CONSIDERATIONS
Quasi-static mobility
1
of surface vehicles depends basically
on vehicle size, traction coefficients (friction) and overall con-
figuration. The Millibot Train has great potential for mobility
due to its relative length and ability to configure itself appro-
priately. With all couplers locked in a roughly straight pose, the
train should be able to cross a ditch nearly half its length. A
standard staircase step is easily crossed by lifting the front end
of the train above the nose of the step, then following an S-curve
to mount the step. The lifting capability is critical for steps with
nose extensions that block directly climbing the vertical face
(riser).
Higher steps, on the order of half the train length, require
more finesse. As shown in Fig. 4(a), the front modules can be
driven straight upward along the face of the step as long as the
rear modules remaining on flat ground have enough power and
traction to push. Simplistically, a train with passive joints can
climb until the weight of the vertical modules is balanced by
the traction forces against the wall. These traction forces de-
pend on the forward thrust generated by the modules on the flat.
The limit is
, where is the friction coefficient (for
both the horizontal and vertical surfaces),
is the total number
of modules in the train, and
is the number of modules against
the vertical face. Thus, for a typical friction coefficient of 0.7,
half of the modules could be driven to vertical; a coefficient of
0.5 allows only one fourth to be lifted. However, the powered
articulation of the Millibot train increases this lifting ability sub-
stantially by augmenting the lifting force applied to the vertical
modules. In reality, the ability to balance the vertical column is
likely to be the limiting factor.
Another limitation on step climbing is the ability to pull the
train onto the plateau at the top of the step. The analysis is sim-
ilar to the previous lifting analysis in that lateral thrust produces
the normal force for tractionagainst the wall. However, therobot
1
“Quasi-static” means that only gravity, contact and traction forces are con-
sidered. Velocities are assumed small so that momentum effects are negligible.
Fig. 5. Prototype used for testing train mobility. Top: detail of a single module,
fabricated largely from FDM parts. Bottom: Seven modules climbing a standard
flight of stairs (left) and a double-height step (right).
must curl around to contact the top surface [Fig. 4(b)], which
places the contact point away from the nose of the step. De-
pending on the friction and geometric parameters, the robot may
be unable to hold this pose and may be stable instead as shown
in pose C.
As with the “pushing” case above, joint actuation can help
to “pull” the robot above the nose of the step, gradually trans-
ferring load to the forward contact point [Fig. 4(d)], providing
traction to drive forward. Clearly, strong joint actuation is crit-
ical for getting over the nose of the step.
IV. M
OBILITY TEST PROTOTYPE
In order to verify the mobility effectiveness of the train con-
cept, a simple, manually-controlled, seven-module prototype
was built as shown in Fig. 5. Each module includes a pair of
small hobby servos independently driving the two tracks. The
tracks were each formed from a pair of thin timing belts (3-mm
wide, 2.03-mm pitch) connected by tubular cross bars glued to
the belts. Thin rubber tubing over each cross bar provided a
traction surface. Each belt was guided on an idler sprocket and
a second sprocket driven by the hobby servo through a short
timing belt. Modules were joined with friction couplers that
would hold the manually set articulation angles. Overall length
of the train was 0.75 m.
An interesting feature of this design was the extensive use
of fused-deposition-modeling (FDM) rapid-prototyping for the
manufacture of most of the parts. Parts were produced from
P1500 polyester on a Stratasys Genisys FDM machine from
ProEngineer
2
CAD models, to a resolution of 0.3 mm. Rela-
tively complex parts, such as the hollow timing-belt sprockets
2
ProEngineer is a registered trademark of Parametric Technology Corpora-
tion, Needham, MA 02494, USA.
BROWN et al.: MILLIBOT TRAINS FOR ENHANCED MOBILITY 455
Fig. 6. Componentsof the module’s track system, including the polycarbonate
driving and idling sprockets (left), the individual track segments (right) and an
assembled track with the polyurethane-covered growsers (rear).
with internal ribs, could be easily and cheaply produced this
way. FDM allowed substantial weight savings compared to con-
ventional machining techniques and enabled quick, low-cost
manufacturing of parts for the seven modules.
The train was wired to a manual switch box with battery pack,
allowing ganged control of the left and right tracks. To emulate
the performance of a train with powered articulations, we alter-
nately drove the machine forward a short distance and manually
adjusted the joint angles. With this method, the train was able to
climb various standard staircases and even a step approximately
0.33-m high.
V. M
ECHANICAL SYSTEMS
Once we had proven the concept of a two-sided, tracked
vehicle in both individual and train configurations, we identi-
fied four major technical hurdles to achieve a fully functional
system: tracks that would grip a variety of surfaces yet require
minimal power to drive; a compact, efficient drive system
to propel the two tracks; a coupling mechanism to allow the
vehicles to join securely into the train configuration and then
separate easily on command; and a strong, compact lifting
mechanism to enable the train to change its shape.
A. Tracks
Our goal for the tracks was to develop a profile that had
large enough growsers (the protrusions sticking out radially
from tank-type treads) to engage the corners of stair treads, but
that ran smoothly on flat surfaces. We wanted tracks that had
low operating friction in order to maximize the available power
and that were securely guided so that they would not be driven
off of their sprockets even when subjected to large side forces.
They also had to be relatively thin to preserve the vehicle’s
small size and wide enough to cover most of the vehicle’s outer
surface to reduce the chance of hang ups.
Although the tracks fabricated for the mobility-test proto-
type provided good traction on a variety of surfaces, the internal
losses in bending the rubber timing belts consumed a large per-
centage of the available drive power. The final, low-loss de-
sign consisted of plastic track segments joined together with
0.8-mm steel pins (Fig. 6). Each segment was injection-molded
in-house from polycarbonate and alternate segments featured
a thin growser onto which polyurethane rubber was cast. Each
segment included a small transverse rib on the underside for lat-
eral alignment. Each track was driven by a round sprocket cut
with axial grooves to engage the track’s hinge protrusions and
cut with a circumferential groove to guide the underside rib.
The resulting design provided for much lower operating
friction, high impact strength and very good traction, owing to
both the growser shape and the sticky polyurethane covering.
In achieving this performance, however, the possibility of
sensors seeing through the track was lost. Manufacturing the
track links was fairly straightforward, although we found that
it was critical to control the mold temperature and flow rate to
prevent nonuniform shrinkage and resultant residual stresses.
The polycarbonate track was tested by running it continu-
ously at operating speed for two weeks while monitoring the
current drawn by the driving motor as a measure of the track’s
operating friction. We found that the friction dropped steadily
during a break-in period of 12 hours and then continued to
run reliably without additional change. Some track segments
showed early cracking and breakage of their hinge tabs, likely
due to residual stresses from the molding process, but such
failures usually occurred within the first six hours of run in,
making it easy to separate out any defective segments from the
final units.
B. Drive
The two tracks are driven by small dc motors through a
speed-reduction mechanism. The engineering challenge was to
develop an efficient, compact, lightweight drive actuator with
sufficient torque for the application. The worst-case design
situation is when the train is climbing tangentially up a flight
of stairs and the power to drive all seven modules is borne by
just four drive units (left and right drives of the two modules
that are touching stair noses). Based on a seven-module train
and estimated module mass of 200 g, the target traction (thrust
force) for each actuator was 1.8 N. To save precious space for
batteries, sensors and electronics, we decided that the entire
drive actuator for each track should fit inside the drive sprocket
for that track. We had hoped to find a commercially-available
motor/drive package that would meet our requirements, but
soon discovered that available dc gearmotors of adequate
power (from MicroMo, Smoovy, Maxon, etc.) were too long
to fit the 2-mm axial space available. In the hobby market,
we found the Mabuchi RF-020TH motor, which fit the space
and provided reasonable torque, but left little axial space for
gear reduction. Based on the available continuous motor torque
(1.5 mN
m from our tests) and expected 75% drive train
efficiency, we needed an approximately 21:1 speed reduction
(torque multiplication) to achieve the desired output torque
with a 26 mm diameter drive sprocket. Corresponding travel
speed under light loading would be about 75 cm/s, ample for
our applications.
The selected motor was 18 mm in length, so some creativity
was needed to achieve the 21:1 speed reduction in less than
7 mm of axial space. We settled on a planetary-traction drive
(Figs. 7 and 8) that enabled the large speed change in a single
stage, as opposed to the two or three stages that planetary
456 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 7, NO. 4, DECEMBER 2002
Fig. 7. Assembly diagram of the planetary-tractiondrive mechanism, showing
the rollers, bearings and shafts (top row), motor/brush assembly (middle), and
drive sprocket with encoder ring and ball bearings (bottom).
Fig. 8. Partially disassembled planetary-traction drive.
gearing would have required. The motor shaft was replaced
with a hardened steel shaft ground to a pinion diameter of 0.9
mm. This drove two overlapping pairs of hard steel rollers
inside a flexible cup of 22 mm inside diameter, giving a drive
ratio of 25:1. The flexible cup, machined from polycarbonate,
provided radial pressure to load the traction surfaces and
served as the drive sprocket for the track. Achieving adequate
pressure for traction without overloading the motor was tricky,
requiring repeated testing and adjustment of cup stiffness,
as well as precise control of part diameters and cup wall
thickness. Also, radial play was designed into the rollers to
allow them to balance the forces between one another without
being unduly constrained by their bearings. The planetary drive
supports the inboard end of each sprocket, while a thin, custom
ball bearing—12 1-mm balls in a plastic cage—supports the
outboard end with minimal running friction.
C. Coupler
The function of the coupler (Figs. 9 and 10) is to securely
lock adjacent Millibot modules together in train mode, while al-
lowing easy engagement and disengagement on command. As
with the other subsystems, strength and compactness were crit-
ical for the coupler. The device needed to fit between the tracks
and sweep out minimal volume in the central cavity (electronics
area) with changing joint angles.
Fig. 9. Cross-sectional view of the coupler used to link together the modules.
Fig. 10. Millibot uses its coupler to link together with another unit.
The design is based on a pair of hard steel pins that register
with a matching receptacle to constrain five degrees-of-freedom
(DOF). The latch wire, shaped from 0.5-mm diameter music
wire, protrudes from the side of one pin to automatically lock
the coupler into the receptacle after two modules have driven
together. Precise parallelism between the two coupler pins and
between the two mating socket holes, plus slight free play (0.03
mm) in the fit permits free insertion and retraction of the pins.
A 150-
m shape-memory-alloy (SMA) wire (Flexinol
150HT) retracts the latch for disengagement when activated
with a heating current of about 0.5 A for one second under
microprocessor control. SMA actuators, although far from
efficient (
10 ), were selected because of their exceptional
specific energy: about 10 J/cm
, an order of magnitude better
than solenoids or other conventional actuators [29]. To accom-
modate the 180
rotation of the joint, a small brass slip-ring and
wiper provide the power connection from the driving circuit
in the central electronics cavity through the SMA wire to the
grounded latch wire.
D. Lifter
The lifting mechanism of the robot presented the toughest de-
sign challenge of any system on the vehicle. Our goal of lifting
three attached modules required a torque of 1.1 N
m and our
available space was only 2-cm diameter by 5-cm long (the space
inside the two idler sprockets). The mechanism had to be non-
backdrivable under normal loads in order to hold its position,
yet protected against damage in case of torque overloads due to
falls, mishandling, etc. Actuation speed was not a concern since
