Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO.

2, APRIL 2020 2913

Magnetic Milli-Robot Swarm Platform: A Safety


Barrier Certificate Enabled, Low-Cost Test Bed
Allen Hsu , Huihua Zhao, Martin Gaudreault, Annjoe Wong Foy, and Ron Pelrine

Abstract—Swarms of micro- and milli-sized robots have the po- in swarm platforms have been limited in size (>16 mm) mainly
tential to advance biological micro-manipulation, micro- assembly due to the size and power constraints (1–24 Hrs) of batteries and
and manufacturing, and provide an ideal platform for studying their method of locomotion: (1) vibratory (i.e. Kilobots [15])
large swarm behaviors and control. Due to their small size and
low cost, tens to hundreds of micro/milli robots can function in and (2) wheeled (i.e. mROBerTO [16], Alice [17], GritsBot [18],
parallel to perform a task that otherwise would be too cumbersome TinyTerp [19], Zooids [20], Jasmine [21]). Other examples of
or costly for a larger macroscopic robot. Here, we demonstrate a milli-robots such as HaMR [25], which use piezo driven legs,
scalable system and modular circuit architecture for controlling have extremely low operational lifetime (<5 minutes). The size
and coordinating the motion of >10’s of magnetic micro/milli of motors, power sources, actuators, and transmissions limit the
robots. By modifying the concepts of safety barrier certificates
to our magnetic robot hardware, we achieve minimally invasive, ultimate size and cost of these robots.
collision-free, 2D position control (x, y) of up to N = 16 robots On the other spectrum of size (μm to mm), many groups
in a low-cost tabletop (288 mm × 288 mm) magnetic milli-robot have used external actuation for locomotion such as magnetic
platform with up to 288 degrees of freedom. We show that the fields [6], [26], [27] or optical forces [28]–[31]. By using ex-
introduction of random dithering can achieve a 100% success rate ternal locomotion the robot itself can be made very simply and
(i.e., no deadlocking), enabling the system to serve as a platform for
the study of various swarm-like behaviors or multi-agent robotic cheaply; however, these systems are primarily limited to aqueous
coordination. environments due to the small locomotive forces and/or are
highly under-actuated due to the size mismatch between the
Index Terms—Micro/nano robots, multi-robot systems, swarms,
Collision avoidance, planning, scheduling and coordination.
robots and the resolution of the external fields [4], [27]. By
enabling more independent controls, asynchronous actions can
I. INTRODUCTION be performed in parallel enabling better fault-tolerant behav-
ior than purely batch-like processes and higher experimental
WARMS of massively parallel micro/milli robots have a
S wide array of potential applications: manipulating biologi-
cal materials [1]–[3], performing medical surgery [4], delivering
throughput.
In our work, we have been investigating a robotic platform
at the meso-scale (μm to mm) to bridge the size and control
drugs [5]–[7], or assembling heterogeneous micro-electronics
gap between aqueous micro-robotic and conventional milli-
and materials [8]–[14]. Furthermore, due to their large numbers,
robot platforms. Using local magnetic fields from a printed
swarms of micro/milli robots can even serve as an experimental
circuit board (PCB) in conjunction with a diamagnetic mate-
platform for studying the science and control methodology of
rial, we have shown multiple independently controlled robots
swarms [15]–[21]. By reducing the robot’s size, cheaper and
(mm-sized) assembling carbon fiber trusses and performing
denser robot populations can be achieved.
micro-assembly [11], [12], [32]–[34]. By leveraging external
Unfortunately, while modern day processors and sensors have
forces from a surface, the robots can be made smaller, cheaper,
continued to shrink, not only in size (<0.04 mm3 ) [22], [23], but
and rapidly deployable, while having large enough forces to
also power consumption (∼500 nW) [24], the locomotion of mi-
transport payloads (∼0.5 g) for carrying additional modules
cro/milli robots at the millimeter and smaller size scales still re-
such as end effectors or electronics. In our past work [34],
main an ongoing scaling challenge. Milli-robots commonly used
>1000 robots could be operated simultaneously; however the
limited scalability of our control hardware restricted the number
Manuscript received September 9, 2019; accepted January 16, 2020. Date of
publication February 18, 2020; date of current version March 1, 2020. This letter of independent actions (N∼18–24). Therefore, in this work we
was recommended for publication by Associate Editor D. Cappelleri and Editor (1) show a novel modular electronic hardware architecture for
X. Liu upon evaluation of the reviewers’ comments. This work was supported by driving up to 288 independent degrees of freedom (DOF) using
SRI International’s internal research and development funding. (Corresponding
author: Allen Hsu.) custom stackable H-bridges to increase the number of indepen-
Allen Hsu, Martin Gaudreault, Annjoe Wong Foy, and Ron Pelrine are with dent actions; (2) introduce closed-loop visual feedback for the
the Advanced Technology and Systems Division, SRI International, Menlo coordination of magnetic coils; and (3) add a low-level collision
Park, CA 94025 USA (e-mail: allen.hsu@sri.com; martin.gaudreault@sri.com;
annjoe.wong-foy@sri.com; ron.pelrine@sri.com). safety guarantee between robots to easily coordinate these new
Huihua Zhao is with Toyota Research Institute, Los Altos, CA 94022 USA found DOF. All three of these components together yield a highly
(e-mail: huihua.zhao@tri.global). versatile, easy-to-maintain and deploy, collision-free motion,
This article has supplementary downloadable material available at http:
//ieeexplore.ieee.org, provided by the authors. low-cost manipulation platform that gives researchers a unique
Digital Object Identifier 10.1109/LRA.2020.2974713 test bed for performing micro/milli manipulation in non-aqueous

2377-3766 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
2914 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 2, APRIL 2020

Fig. 2. Zone Command Generation. (a) Diagram of how low-level zone com-
mands are formed based on the position of the robot. As the robot moves from
one zone to another, additional zones must be activated. v represents the velocity
of the robot. Note the robot’s orientation is constant as it moves. (b) Diagram of
constructing global zone command from the individual zone commands of each
robot by summing the individual commands and then normalizing each zone by
the number of duplicate commands (nZ ).

while the orientation of the robot is constant when driven. In


past works [32], [33], we have shown in greater detail how an
analog waveform can generate sub-micron continuous motion
when the robot is diamagnetically levitated; however, digital
stepping is used here to simplify the required electronics. By
Fig. 1. Magnetic Milli-Robot Theory. (a) Schematic and picture of magnetic
robot and serpentine traces (Top and Side View) used to generate locomotion
using multi-pole magnetic arrays, robots can operate at 1–2 body
for independent zones. I1 , I2 and I3 , I4 control x and y motion, respectively. length’s away (i.e. 12–14 mm [34]) due to the field cancellation
Scale bar represents 5 mm. Robot shown with AprilTag attached. (b) Diagram from alternating magnetic dipoles.
of quadrature current switching as a function of time for 1D step motion of the
robot. I1 , I2 are stepped through a sequence of (+, +), (−, +), (−, −), (+, −).
Blue and red arrows illustrate magnetic field lines due to currents in the traces. B. Low Level Multi-Zone Coordination
(c) Diagram of coordination among independently controlled zones. Only when
both Zone 1 and Zone 2 have the same polarity and direction is there no magnetic To independently control multiple robots, we can tile our
discontinuity on the board.
serpentine traces in space to form a zone of control (24 mm ×
24 mm). This size was selected based on commercially available
electrical headers used to make contact with the PCB. A single
environments, as well as, a potential table-top test bed for safely zone can move all robots within its bounds; however, their
studying the control of large numbers of interacting robots. motions are under-actuated (i.e. the robots’ motions within a
zone are no longer independent from each other). Conversely,
II. SYSTEM DESCRIPTION due to the magnetic potential generated by our serpentine
traces, controlling multiple robots operating across multiple
A. Magnetic Robot Theory
zones requires careful coordination. In Fig. 1c, Zone 1 is in
The milli-robots described in this work consist of mul- a configuration of (I1 |I2 )Z1 = (+|+). The robot can pass from
tipole magnetic arrays (i.e., alternating north-south checker- Zone 1 to Zone 2 only if Zone 2 is in the same configuration
boards) of NdFeB magnets (N52 M, individual magnet size: (I1 |I2 )Z2 = (+, +). Otherwise, the magnetic potential barrier
1.4 mm × 1.4 mm × 0.4 mm, full robot size: 5 × 5 magnetic ar- has a discontinuity and is not periodic at the boundaries between
rays: 7 mm × 7 mm). The PCB consists of four layers of serpen- zones. Furthermore, like a conveyor belt, both zones must be
tine traces providing motion of a robot in the X and Y direction moving in the same direction to transfer the robot successfully
(Fig. 1a). The serpentine coil/drive pattern generates a periodic from one zone to the next. To simplify the notation, we represent
magnetic potential that locks the robot to discrete positions with the motion for each of the zones as a zone command array
a constant orientation. Fig. 1b shows how step motion is achieved (Zcmd ) = [Z1X , Z1Y , . . ..ZN X , ZN Y ], where ZN X and ZN Y
(0.5 mm). Stepping through a sequence of currents in forward or represents a discrete distance moved in integer multiples of
reverse order (i.e. [(I1 |I2 )] = [(+|+), (−|+), (−|−), (−|+)]) 0.5 mm, which is the smallest discrete step size defined by the
generates a forward or reverse traveling surface magnetic wave serpentine traces.
in a single direction, confined within 1–2 mm away from the Fig. 2a illustrates how multi-zone coordination works depend-
board surface (< 100 μT ). Due to the trace’s symmetry, the same ing on the robot’s position: (i) When the robot is alone within a
also applies for the orthogonal direction. The forces generated zone, only a single zone must be addressed. (ii) When the robot
by the traces can translate the center of mass of the robot, enters the transition region between two zones (shown in pink),

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
HSU et al.: MAGNETIC MILLI-ROBOT SWARM PLATFORM: A SAFETY BARRIER CERTIFICATE ENABLED, LOW-COST TEST BED 2915

both zones must be of the same phase and moving in the same
direction until the entire robot remains solely within a single
zone. The transition region is set to the diameter of the robot
plus an additional safety margin (∼ 4 mm) due to the positional
uncertainty from feedback. (iii) In 2D, we must also worry about
the situation where corners of four zones meet.
Coordination of multiple robots and zones is slightly more
complicated. Fig. 2b shows the coordination of a robot formation
(i.e., when the inter-robot distances are constant). This can still
be treated as a summation of the individual zone commands,
normalized by the number of duplicate commands per zone (nZ ).
Fig. 3. Safety Barrier Certificates. (a) Two robots, if they continue with their
Given the position and trajectory of all robots on the board, original velocity û, will eventually have an inter-robot distance less than their
low-level coil commands or current sequences can be generated defined safety radius rsaf ety . By solving Eq. (1), a new velocity u∗ is computed
for the entire board. and then executed by the robots. (b) Due to the hardware drive of the system, u∗
must be discretized in both space and time u∗d in order to map to the low-level
zone commands. The robot can move in discrete directions as indicated by the
grey arrows.
C. Collision Avoidance Controller
To coordinate the robots’ motion, we implemented a low-level
collision avoidance controller that serves as the foundation for with
other higher-level multi-agent path planning and scheduling al-  
Aij = 0, . . ., −2(pi − pj )T , . . ., 2(pi − pj )T , . . ., 0 , (4)
gorithms. There exist two main categories of collision avoidance
methods found in the literature: reactive (reciprocal velocity hij (x) = pi − p̂j 2 −rsaf
2
ety , (5)
obstacles [35], potential fields [36]), and optimization based
where γ is positive and umax is the maximum allowed com-
methods (model predictive control [37], mixed integer linear
manded velocities.
programming [38] safety barrier certificates (SBCs) [39], [40]).
While reactive algorithms are computationally fast, they cannot
D. Discretization Process
forecast future states of neighboring robots. Conversely, opti-
mization based methods utilize reactive algorithms subject to To apply the controller output in Eq. (1) to our magnetic
constraints (i.e. future time evolution of robots or minimal inva- robots, we must compute the discrete optimal control input(u∗d ),
siveness) to compute the output velocity. In this work, we elected by discretizing u∗ = (u∗x , u∗y ) with respect to the direction
to use the formalism of safety barrier certificates (SBCs) for the θ∗ = tan−1 (u∗y /u∗x ). Eqs. (6) and (7) describe the discretization
following two reasons: First, of the optimization based methods, process, where k is an integer determining how many steps to
SBCs provides formal guarantees about collision avoidance by discretize θ∗ , and v0 is the minimum discrete step the robot
assigning a safety radius (rsaf ety ) [39], [40], which, in our case, can take. In this work,we set k = 2 and k = 4, which limits
ensures the magnetic force between robots is always less than the robots to discrete steps of either 90-degree motions, or
the forces generated by the underlying traces and any static 45-degree motion resolution (Fig. 3b). In addition, we also
friction due to the robot’s weight. Second, past work has shown introduce a discrete random variable, χ, which can take on the
that SBCs are a scalable solution toward decentralized collision values {−1, 0, 1} with equal probability. Because the robot’s
avoidance for hundreds of robots [39], [41]. underlying motion is discrete in time and space, deadlocking
In particular, we model each robot i as a single integrator can occur because physical space is a shared resource (e.g. two
with dynamics ṗi = ui , where pi ∈ R2 and ui ∈ R2 represent robots are about to intersect, so they back away at the same time,
the positions, and control inputs (i.e., velocities) respectively. but then they both try to re-enter the same space again, only
Given an initial control input (û) from any user provided nominal to repeat the same cycle indefinitely). Therefore, introducing
controller, the SBCs computes a new control input (u∗ ) that a random dither to the actual control velocity can avoid such
provides three guarantees a) the outputs are collision-free, b) problems between robots since on the next cycle, they will both
minimum modification to the original input only if the original no longer take the same exact action.
control input leads to a collision, c) the outputs are within the  
π θ∗ π
hardware limits [42]. Fig. 3a shows the concept of this procedure. θd = +χ (6)
k π/k k
We briefly show the formulation for SBCs; however, we refer
the readers to [39], [42], [43] for additional details. u∗d = (v0 cos θd , v0 sin θd ) (7)


N III. METHODS
u∗ = argmin  ui − ûi 2 (1)
u∈R21 i=1 A. Hardware
s.t. Aij u ≤ γhij (x) ∀i = j (2) Fig. 4a shows a picture of our entire tabletop magnetic swarm
platform, with a PCB drive surface, a 12 × 12 matrix of 144
 ui  ≤ umax (3) individually addressable zones (i.e., 4 serpentine trace patterns

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
2916 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 2, APRIL 2020

Fig. 4. Magnetic Milli-Robot Swarm Platform. (a) Picture of the entire tabletop platform (288 mm × 288 mm), including all switching electronics and micro-
controller. Not shown in the picture is a top-down camera for robot tracking. (b) Modular digital magnetic drive unit that drives a line of zones in the magnetic drive
surface. Each drive unit contains its own 4-channel current source with daisy-chained H-bridges. (c) Circuit schematic of modular digital magnetic drive unit. To
simplify the gate drive voltage, Zener diodes (not shown in the diagram) are placed between the gate and source of each MOSFET. A single gate voltage can be
used to drive all the H-bridges, which are ultimately controlled by a 5 V signal from a series of daisy-chained shift registers.

with 2 DOF each), totaling up to 288 DOF. Fig. 4a shows the efficient current source. We inserted into the feedback loop
PCB without any diamagnetic pyrolytic graphite surface, which of the buck circuit, a digital potentiometer to also enable
can be used to decrease robot-PCB friction and increase robot dynamic power scaling and tuning of the drive currents in
reliability [11]. the board, which are nominally driven at (I1 , I2 , I3 , I4 ) =
1) Stackable H-Bridges: Since traces within each zone are (0.25 A, 0.3 A, 0.50 A, 0.7 A). A standard 5 V digital signal is
identical, using a single current source and arranging each zone used as an input for the gate drivers for the H bridge. The polarity
in series with one another can greatly reduce the number of of each H-bridge is stored in an array of daisy-chained shift
electronic components (such as cables and connectors). Fig. 4b registers, whose state is populated by an Arduino MEGA 2560
and 4c show our stackable modular H-bridge. While single microcontroller with an adjustable refresh rate of 100 − 200 Hz
H-bridges are relatively common components, daisy-chaining for the entire board. The refresh rate determines the maximum
H-bridges is more challenging due to the need for complex high- rate of change in current and consequently the robots’ velocities.
side gate drivers since the transistor source voltage is floating. To 3) Full System: Fig. 4b shows an individual modular digital
overcome this challenge, we utilize Zener diodes placed between magnetic drive unit, which consists of 12 H-bridges stacked in
the gates and sources of all the MOSFETS in the H-bridge to series and a single 4× current source. Twelve of these units
clamp the VGS of all transistors to a maximum voltage. There- power the entire magnetic drive surface. To interface and con-
fore, all transistors within a stack can be driven by a single gate trol the system, we use a workstation running Ubuntu 16.04
driver voltage (VGAT E ), such that VGAT E + VT H > VM AX , communicating to the microcontroller over a Universal Serial
where VM AX is the voltage at the output of the current source, Bus (USB) COM port. We used a 5.0 megapixel (MP), FLIR
and VT H is the turn-on voltage for the MOSFET in the H-bridge. Blackfly S Camera (with Sony IMX250 2/3 ) with a Fujinon
2) Current Source and Shift Registers: To power the 4 traces HF8XA-5 M 2/3 8.3 mm 5 MP Machine Vision Lens for
within each serpentine path, we used four buck converters performing visual position feedback for the robots. Up to 144
(LM2673 from Texas Instruments), each serving as a highly micro-robots can be independently controlled on this PCB, while

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
HSU et al.: MAGNETIC MILLI-ROBOT SWARM PLATFORM: A SAFETY BARRIER CERTIFICATE ENABLED, LOW-COST TEST BED 2917

Fig. 5. ROS Software Architecture. (1) Machine Vision for camera calibration
and robot tracking, (2) Robot Controller including safety barrier certificates and
low-level driver to produce position feedback zone commands. Compiled zone
commands are sent over serial COM to the microcontroller.

even more (N ∼ 300) can be controlled dependently (common


motion), depending on the robot size [34]. The total cost of the
system including all of the electronics is ∼ $1500.
Fig. 6. Magnetic Milli-Robot System Simulator. (a) ROS diagram of hardware
simulator software architecture. (b) Simplified physics simulation of robot’s
B. Software motion, by computing the total forces acting on the center of mass of the
robot (Eq. (11)). For simplicity, torques are ignored, but a static friction term is
The Robot Operating System (ROS) is used to combine and included. The forces generated on the robot are normalized forces. (c) Simulated
trajectories for robots based on start conditions and end goal position provided
link various messages between software nodes [44]. The ROS by the Robot Scheduler.
pipeline is divided into three main modules: Machine Vision,
Robot Controller, and Hardware Interface (shown in Fig 5).
1) Machine Vision: We use fiducial tags for both robot track-
normal distance between the midpoint xM and the line seg-
ing and automated camera calibration [45], specifically the
ment formed by two endpoints x1 and x2 as Eq. (10). After
AprilTag family 36h11 which contains 586 distinct IDs. Using
removing the barrel distortion, we calculate the homography
the AprilTags3 library applied to a 2448 × 2048 image with
matrix to transform between pixel coordinates and real spatial
2× decimation, we can detect a 7 mm × 7 mm AprilTag within
coordinates.
the field of view (∼ 300 mm × 300 mm) at approximately 6.43
2) Robot Controller: To control the robots’ position, a Robot
frames per second (fps) through parallel multiprocessing of
Scheduler generates a list of robot goal positions and eval-
camera frames over the entire 5 MP image.
uates whether a robot has reached its target. We utilize a
To accurately estimate the robot’s position, we implement
nominal position based proportional-integral-derivative (PID)
an auto-calibration routine to compute the transformation from
controller to generate (û), which computes a straight-line tra-
the camera‘s coordinate system to the magnetic drive surface’s
jectory between the current and target position. û is then
reference frame. Included in this routine is a method to remove
passed to the SBC, which then modifies û as necessary, to
barrel distortion found near the corners and edges of the image.
avoid any robot collisions. All the robot commands are then
AprilTags are placed along the edges of the magnetic drive
consolidated by the driver as discussed in Section IIB to
surface near the limits of the camera’s field of view (white
generate a single unified zone command that is then sent to
border in Fig. 4a). Calibration tags along each side of the image
the hardware interface ROS node, which relays these com-
should be collinear; however, due to barrel distortion, they form
mands to the microcontroller. While the robot controller can
a parabola. The barrel distortion parameter (κ) is chosen to
run up to 1 kHz, the overall update rate of zone commands
minimize the total parabolicity of all four lines:
is limited by the feedback refresh rate of the machine vision
3
 module.
min parabolicity(x1,i , x2,i , xM,i ) (8)
κ
i=1 IV. EXPERIMENTS
xpolar (r, θ) −−−−−−−−→ xpolar (r(1 + κr2 ))
barrel distortion
(9) A. Simulation
(x2 − x1 )(x1 − xM ) To optimize our SBC controller, we implemented a simple
parabolicity(x1 , x2 , xM ) = (10) magnetic robot simulator in ROS (Fig. 6a). We model the robot
x2 − x1 
as a simple point object and compute the total force applied to
where (r, θ) and x are polar coordinate representations of the robot, while neglecting torque. For computational reasons,
the pixel positions and the pixel coordinate position, respec- we simplify the physics as seen by the robot (Fig. 6b). The total
tively. We define the parabolicity of a line segment as the force (Ftot ) is the weighted area (A) sum of all forces applied

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
2918 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 2, APRIL 2020

 ) to the robot:
(dF


Nz
Ftot =  i)
c(Ai dF (11)
i=0

where Nz is the total number of zones that overlap the robot


and c is a normalization constant. We also include static friction
(Fμ ), by specifying that Ftot  > Fμ  for the robot to move.
To simulate possible motion errors for each time step, we can
multiply Ftot by a binary random variable with values of {0, 1}
and with a probability of motion that depends on a sigmoidal
distribution defined by Ftot  − Fμ . Fig. 6c shows the sim-
ulated trajectories for five robots without a random dithering
term.
To test our SBC controller implementation and tune controller
parameters, we used two performance metrics: normalized dis-
tance traveled (i.e. total distance traveled/ ideal straight-line
distance = Lact /Lideal ) and success rate, η (% trials where all
robots reached their final objectives without deadlocking). To
avoid magnetic dipole forces and the situation of under-actuated
control, we set rsaf ety = 34 mm, which is chosen based on the Fig. 7. Simulated Trajectories for Varying Number of Robots. Initial start
configuration is a random location of robots with the caveat that no two robots
diagonal length of a zone and an additional safety margin. The violate the safety radius constraint. In the target goal position, robots are spaced
barrier gain was tuned based on trial and error. We studied two uniformly around a circle. Shown above and below are N = 5 robots and N =
parameters of our SBC controller: number of robots (N ) and 20 robots, respectively; their corresponding trajectories are color-coded.
stochastic/deterministic robot motion (i.e., with and without the
dithering component χ in Eq. (6). N robots were randomly
placed across the board, such that no robots violated the safety
distance constraint. We repeated this for 30 trials, alternating
between a random configuration and a uniformly spaced circle.
Fig. 7 shows the simulated path of robots using the SBC con-
troller, including a random dithering term. In contrast to Fig. 6c,
the robots’ trajectories with dithering approximate more of a
random walk. Initially all robots attempt to take a straight-line
approach to their final objective. As their paths begin to intersect,
the robots adjust their trajectory to avoid collisions, causing them
to naturally circle one another. Increased interactions between
robots (which depends on the initial position) and random dither-
ing increases the average distance traveled for all the robots Fig. 8. Histograms of Simulated Trajectories. X-axis is the distance traveled
by every robot (Lact) normalized to their ideal straight-line trajectory (Lideal).
(Fig. 8); however, the stochastic term increases the success rate Graphs on the left and right are for N = 5 and N = 20 robots, respectively.
of robots reaching their destination by preventing deadlocking Graphs on the top and bottom are for trajectories with and without random dither
between robots, due to velocity discretization. As the number (stochastic versus deterministic discretization). The mean distance traveled and
success rate η are labeled in each graph.
of robots increases, the distribution of path lengths begins to
approximate a Maxwell-Boltzmann distribution as the robots
begin to resemble an ideal gas.
any pre-computed routes. We conducted experiments for k = 4;
however, the robot motion was less robust due to torques on
B. Hardware
the robot [46]. We observed η = 100%, neglecting issues with
To validate the entire system and hardware controller, we the robot, and a normalized path length of Lact /Lideal = 3.16 ±
ran similar experiments on the actual hardware (Fig. 9). For 3.07 for 18 trials. The average time per trial was 74.13 ± 30.13 s
illustrative purposes, we replaced the formations from a random with a robot velocity of approximately 8–9 mm/s, which is
initial start point to a sequence of repeating alpha-numeric set by the refresh rate of the position feedback. To increase
letters of “S,” “R,” “I”. Fig. 9a-c show images from the ROS the locomotion’s robustness, higher-order corrections such as
software of the end configuration of 16 robots. Fig. 9d and 9e torque should be included [46], since robots interacting with
show the measured trajectories from hardware as they transition multiple zones can experience applied torques. Modifications to
from one configuration to the next for k = 2 (i.e. 0/90-degree our existing controller (i.e. only dithering when the robots begin
directions) as computed in real-time by the controller without to interact or when deadlocking is detected) or implementing

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
HSU et al.: MAGNETIC MILLI-ROBOT SWARM PLATFORM: A SAFETY BARRIER CERTIFICATE ENABLED, LOW-COST TEST BED 2919

TABLE I
ROBOT COMPARISON TABLE

Fig. 9. Hardware Results of Magnetic Milli-Robot Pattern Formation.


(a) Camera image with robots tracked in blue. N = 16 robots start in the letter
“S”., then (b) “R,” and then (c) “I”. This sequence then loops back on itself and is
repeated 6 more times. Charts in (d) and (e) show the plotted robot path from ROS
data of all robots. Final goal position is represented as a solid- colored diamond.
While the final goal position of the robots is the same, the individual locations
are randomized among the different unique robot IDs for every formation.

other low-level collision avoidance algorithms [35]–[38] may


composed of low-cost commercially available components and
further improve η and the path length of the robot.
generates sufficient forces to operate in ambient conditions and
can hold macro-like payloads.
V. CONCLUSION Finally, while the platform has the potential for performing
Using modular electronics in conjunction with SBCs, we swarm studies, the system also serves as a useful hardware plat-
have demonstrated a tabletop magnetic milli-robot swarm form to support the distributed and collaborative manufacturing
platform with up to 288 DOF and collision-free motion of by swarms of milli-robots [11], [34]. Past work has already
16 robots within the shared space. Assuming we are limited by shown that simple capillary and dry-adhesive end effectors can
attractive inter-robot forces (12–15 mm) this implies N ∼ 180 be added to these simple robot bases to assemble carbon fiber
can operate within the 288 mm × 288 mm platform. While rods, liquid adhesives, surface mount electronic components,
many of the critical dimensions such as rsaf ety (39 mm) and fiberglass sheets, and micro-components [11], [34]. Continued
zone size (24 mm) are larger than the robot’s size, these were advancements in miniaturization of onboard sensors and actu-
chosen due to the ease of PCB and magnet manufacturing. ators from the micro-robotics community can serve to further
Currently the lack of high resolution magnetic patterning enhance the capabilities and dexterity of our magnetic milli-
methods of NdFeB ultimately limit the smallest achievable robots. In addition, by using analog waveforms and diamagnetic
multi-pole array, which is required for our robots. levitation, these high motion precision milli-robots can serve as
Table I compares the hardware capabilities of this system as a massively parallel platform for interfacing between existing
compared to other similar sized swarm milli-robots platforms micro- and macro-robots. This work’s advancements in collision
and externally actuated micro/milli robots. In this work, using avoidance, distributed formation control, and large degrees of
externally actuated local magnetic fields, the system can operate freedom can now enable truly distributed and collaborative
indefinitely, achieve smaller (2–7 mm) robot sizes [34], and manufacturing where robots co-located can interact with each
lower cost ($5–15 per robot) than larger sized milli-robots. While other to form mini-work cells or where distributed control and
the robots in this work contain no local processing or direct scheduling algorithms can be examined [47].
communication, the robots can still locally interact based on
their relative distances as seen by the external camera similar to REFERENCES
Zooids [20]; moreover, the payload of the robot scales roughly
with the in-plane area of the robot; therefore, the weight of a [1] M. Sitti et al.,“ Biomedical applications of untethered mobile
milli/microrobots,” Proc. IEEE, vol. 103, no. 2, pp. 205–224, Feb. 2015.
4.8 mm coin-cell battery and onboard processor/transmitter [22], [2] F. Arai, A. Kawaji, P. Luangjarmekom, T. Fukuda, and K. Itoigawa, “Three-
[23] can still be easily carried by a robot of only ∼ 3 mm × dimensional bio-micromanipulation under the microscope,” in Proc. IEEE
3 mm. Relative to other micro-robotic platforms, the system is Int. Conf. Robot. Autom. (Cat. No.01CH37164), 2001, vol. 1, pp. 604–609.

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.
2920 IEEE ROBOTICS AND AUTOMATION LETTERS, VOL. 5, NO. 2, APRIL 2020

[3] E. E. Hunter, E. B. Steager, A. Hsu, A. Wong-Foy, R. Pelrine, and V. [25] B. Goldberg et al., “Power and control autonomy for high-speed locomo-
Kumar, “Nanoliter fluid handling for microbiology via levitated magnetic tion with an insect-scale legged robot,” IEEE Robot. Autom. Lett., vol. 3,
microrobots,” IEEE Robot. Autom. Lett., vol. 4, no. 2, pp. 997–1004, no. 2, pp. 987–993, Apr. 2018.
Apr. 2019. [26] S. Chowdhury, W. Jing, and D. J. Cappelleri, “Towards independent control
[4] M. P. Kummer, J. J. Abbott, B. E. Kratochvil, R. Borer, A. Sengul, and of multiple magnetic mobile microrobots,” Micromachines, vol. 7, no. 1,
B. J. Nelson, “OctoMag: An electromagnetic system for 5-DOF wireless pp. 3–17, Dec. 2015.
micromanipulation,” IEEE Trans. Robot., vol. 26, no. 6, pp. 1006–1017, [27] E. Diller, J. Giltinan, and M. Sitti, “Independent control of multiple
Dec. 2010. magnetic microrobots in three dimensions,” Int. J. Robot. Res., vol. 32,
[5] O. Felfoul et al., “Magneto-aerotactic bacteria deliver drug-containing no. 5, pp. 614–631, Apr. 2013.
nanoliposomes to tumour hypoxic regions,” Nature Nanotechnol., vol. 11, [28] J. E. Curtis, B. A. Koss, and D. G. Grier, “Dynamic holographic optical
no. 11, pp. 941–947, Nov. 2016. tweezers,” Opt. Commun., vol. 207, no. 1, pp. 169–175, Jun. 2002.
[6] S. Schuerle et al., “Robotically controlled microprey to resolve initial [29] F. Arai, K. Onda, R. Iitsuka, and H. Maruyama, “Multi-beam laser micro-
attack modes preceding phagocytosis,” Sci. Robot., vol. 2, no. 2, Jan. 2017, manipulation of microtool by integrated optical tweezers,” in Proc. IEEE
Art. no. eaah6094. Int. Conf. Robot. Autom., May 2009, pp. 1832–1837.
[7] G. Hwang et al., “Catalytic antimicrobial robots for biofilm eradication,” [30] P. Y. Chiou, A. T. Ohta, and M. C. Wu, “Massively parallel manipulation
Sci. Robot., vol. 4, no. 29, Apr. 2019, Art. no. eaaw2388. of single cells and microparticles using optical images,” Nature, vol. 436,
[8] J. Agnus et al., “Robotic microassembly and micromanipulation at no. 7049, pp. 370–372, Jul. 2005.
FEMTO-ST,” J. Micro-Bio Robot., vol. 8, no. 2, pp. 91–106, Apr. 2013. [31] M. A. Rahman, J. Cheng, Z. Wang, and A. T. Ohta, “Coopera-
[9] H. Aoyama and O. Fuchiwaki, “Flexible micro-processing by multiple tive micromanipulation using the independent actuation of fifty mi-
microrobots in SEM,” in Proc. IEEE Int. Conf. Robot. Autom., 2001, vol. 4, crorobots in parallel,” Scientific Rep., vol. 7, no. 1, pp. 1–11, Jun.
pp. 3429–3434. 2017.
[10] D. Cappelleri, D. Efthymiou, A. Goswami, N. Vitoroulis, and M. Zavlanos, [32] R. Pelrine et al., “Diamagnetically levitated robots: An approach to mas-
“Towards mobile microrobot swarms for additive micromanufacturing,” sively parallel robotic systems with unusual motion properties,” in Proc.
Int. J. Adv. Robot. Syst., vol. 11, no. 9, pp. 150–164, Sep. 2014. IEEE Int. Conf. Robot. Autom., May 2012, pp. 739–744.
[11] A. Hsu et al., “Application of micro-robots for building carbon fiber [33] A. Hsu et al., “Diamagnetically levitated milli-robots for heterogeneous
trusses,” in Proc. Int. Conf. Manipulation, Autom. Robot. Small Scales, 3D assembly,” J. Micro-Bio Robot., vol. 14, no. 1, pp. 1–16, Jun. 2018.
Jul. 2016, pp. 1–6. [34] SRI International, “Magnetically actuated micro-robots for advanced ma-
[12] A. Hsu et al., “Automated 2D micro-assembly using diamagnetically nipulation applications,” YouTube, Apr. 9, 2014, [video file]. [Online].
levitated milli-robots,” in Proc. Int. Conf. Manipulation, Autom. Robot. Available: https://www.youtube.com/watch?v=uL6e3co4Qqc. Accessed
Small Scales, Jul. 2017, pp. 1–6. on: Feb. 24, 2020.
[13] B. R. Donald, C. G. Levey, and I. Paprotny, “Planar microassembly [35] J. van den Berg, M. Lin, and D. Manocha, “Reciprocal velocity obstacles
by parallel actuation of MEMS microrobots,” J. Microelectromechanical for real-time multi-agent navigation,” in Proc. IEEE Int. Conf. Robot.
Syst., vol. 17, no. 4, pp. 789–808, Aug. 2008. Autom., May 2008, pp. 1928–1935.
[14] Y. Liu and N. M. Ravindra, “A magnetic-field-assisted milli-scale robotic [36] E. Rimon and D. Koditschek, “Exact robot navigation using artificial
assembly machine: An approach to parallel robotic automation systems,” potential functions,” IEEE Trans. Robot. Autom., vol. 8, no. 5, pp. 501–518,
Micromachines, vol. 9, no. 4, pp. 144–157, Mar. 2018. Oct. 1992.
[15] M. Rubenstein, C. Ahler, and R. Nagpal, “Kilobot: A low cost scalable [37] W. Dunbar and R. Murray, “Model predictive control of coordinated multi-
robot system for collective behaviors,” in Proc. IEEE Int. Conf. Robot. vehicle formations,” in Proc. 41st IEEE Conf. Decis. Control, Dec. 2002,
Autom., May 2012, pp. 3293–3298. vol. 4, pp. 4631–4636.
[16] J. Y. Kim, T. Colaco, Z. Kashino, G. Nejat, and B. Benhabib, “mROBerTO: [38] T. Schouwenaars, B. De Moor, E. Feron, and J. How, “Mixed integer
A modular millirobot for swarm-behavior studies,” in Proc. IEEE/RSJ Int. programming for multi-vehicle path planning,” in Proc. Eur. Control Conf.,
Conf. Intell. Robots Syst., Oct. 2016, pp. 2109–2114. Sep. 2001, pp. 2603–2608.
[17] G. Caprari and R. Siegwart, “Mobile micro-robots ready to use: Al- [39] L. Wang, A. D. Ames, and M. Egerstedt, “Safety barrier certificates for
ice,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Aug. 2005, collisions-free multirobot systems,” IEEE Trans. Robot., vol. 33, no. 3,
pp. 3295–3300. pp. 661–674, Jun. 2017.
[18] D. Pickem, M. Lee, and M. Egerstedt, “The GRITSBot in its natural habitat [40] U. Borrmann, L. Wang, A. D. Ames, and M. Egerstedt, “Control barrier
- A multi-robot testbed,” in Proc. IEEE Int. Conf. Robot. Autom., May 2015, certificates for safe swarm behavior,” IFAC-PapersOnLine, vol. 48, no. 27,
pp. 4062–4067. pp. 68–73, Jan. 2015.
[19] A. P. Sabelhaus, D. Mirsky, L. M. Hill, N. C. Martins, and S. Bergbreiter, [41] D. Pickem et al., “The Robotarium: A remotely accessible swarm robotics
“TinyTeRP: A tiny terrestrial robotic platform with modular sensing,” in research testbed,” in Proc. IEEE Int. Conf. Robot. Autom., Singapore,
Proc. IEEE Int. Conf. Robot. Autom., May 2013, pp. 2600–2605. May 2017, pp. 1699–1706.
[20] M. Le Goc, L. H. Kim, A. Parsaei, J.-D. Fekete, P. Dragicevic, and S. [42] A. D. Ames, X. Xu, J. W. Grizzle, and P. Tabuada, “Control barrier function
Follmer, “Zooids: Building blocks for swarm user interfaces,” in Proc. based quadratic programs for safety critical systems,” IEEE Trans. Autom.
29th Annu. Symp. User Interface Softw. Technol., New York, NY, USA, Control, vol. 62, no. 8, pp. 3861–3876, Aug. 2017.
2016, pp. 97–109. [43] X. Xu, P. Tabuada, J. W. Grizzle, and A. D. Ames, “Robustness of control
[21] S. Kornienko and S. Kornienko, “IR-based communication and perception barrier functions for safety critical control” IFAC-PapersOnLine, vol. 48,
in microrobotic swarms,” Sep. 2011, arxiv.org/abs/1109.3617. no. 27, pp. 54–61, Jan. 2015.
[22] F. Maksimovic et al., “A crystal-free single-chip micro mote with inte- [44] M. Quigley et al., “ROS: An open-source robot operating system,” in Proc.
grated 802.15.4 compatible transceiver, sub-mW BLE compatible beacon ICRA Workshop Open Source Softw., vol. 3, Jan. 2009, pp. 1–6.
transmitter, and cortex M0,” in Proc. Symp. VLSI Circuits, Jun. 2019, [45] E. Olson, “AprilTag: A robust and flexible visual fiducial system,” in Proc.
pp. C88–C89. IEEE Int. Conf. Robot. Autom, May 2011, pp. 3400–3407.
[23] X. Wu et al., “A 0.04 mm316 nW wireless and batteryless sensor system [46] R. Pelrine, A. Hsu, and A. Wong-Foy, “Methods and results for rotation
with integrated cortex-M0+ processor and optical communication for of diamagnetic robots using translational designs,” in Proc. Int. Conf.
cellular temperature measurement,” in Proc. IEEE Symp. VLSI Circuits, Manipulation, Autom. Robot. Small Scales, Helsinki, Finland, Jul. 2019,
Jun. 2018, pp. 191–192 pp. 1–5.
[24] Z. Qian, S. Kang, V. Rajaram, C. Cassella, N. E. McGruer, and M. [47] Y. Kantaros, B. V. Johnson, S. Chowdhury, D. J. Cappelleri, and M. M.
Rinaldi, “Zero-power infrared digitizers based on plasmonically enhanced Zavlanos, “Control of magnetic microrobot teams for temporal micro-
micromechanical photoswitches,” Nature Nanotechnol., vol. 12, no. 10, manipulation tasks,” IEEE Trans. Robot., vol. 34, no. 6, pp. 1472–1489,
pp. 969–973, Oct. 2017. Dec. 2018.

Authorized licensed use limited to: University of Exeter. Downloaded on May 08,2020 at 08:04:42 UTC from IEEE Xplore. Restrictions apply.

You might also like