Building a Passive Robot for Active Learning

A Junior Independent Study Self-Design

Justine Walker
Physics Department, The College of Wooster, Wooster, Ohio 44691, USA
(Dated: May 5, 2017)
We found that creating a passive walking robot out of plastic Tinkertoys is possible. This toy
is able to walk down a slope without a motor, simply by the force of gravity. To understand this
system, we videotaped multiple runs of the walker taking multiple steps down an incline. Then we
used motion tracking and video-editing software to measure the length in time and distance of each
step. By plotting these measurements against the Cornell group that inspired this project’s average
measurements, we discovered that the differences in the mass distribution for a wooden and plastic
Tinkertoy walker change the behavior of the system. We found that the time step was faster in our
system (0.34 ± 0.04 s compared to 0.47 s) and the step distance was longer in our system (1.8 ± 0.3
cm compared to 1.3 cm). The percent difference between our group and the Cornell group for both
of these values was 32%. These constrained double pendulum systems are extremely sensitive to the
building parameters and the initial conditions. Because of this, this system is hard to accurately
predict using simplified models.

I. INTRODUCTION

The field of robotics is hot right now across the STEM

world. The applications are seemingly endless. When
people think of robotics they usually think of metallic,
motor-driven machines. That does an injustice to the
expansive possibilities of robots, including soft and pas-
sive robotics.
Passive robotics is an important subfield for multiple
reasons. If we learn how to efficiently use energy that is
already a part of a system we can create robots that are
more energy efficient. For our study passive robotics are
significant because they teach us why certain movements
are efficient and prevalent in our world.
In our work we looked at passive walkers, which are FIG. 1. Diagram of our passive walker simulation’s variables.
walkers that are able to move down a slope simply by This diagram helps clarify the meaning of key terms, such as
the force of gravity driving them down it. This kind swing leg and stance leg. It also denotes which angle is which
of observed movement teaches us why humans have a in our theory. The picture specifies what vectors and angles
tendency walk on earth. Since the simple walkers we correspond with walker before and after the jump conditions
model and create are able to move down a slope without are applied. The first five parts of the diagram represent the
a driving force it means that walking is extremely efficient movement of a complete step before a jump, while the last
in earth’s gravity field. This reason is why humans are part shows the most directly after the jump occurs, when the
able to walk for such long distances without getting tired. walker puts the swing foot down and Equation 10 is satisfied.
We modified a diagram from Mariano Garcia’s Thesis to have
Other gaits, such as running or hopping, are less efficient
the specifications of our simulation. [4]. For our work we
under earth’s gravity because they require more inputted included angles to simplify our solution (i.e. the alpha angles)
energy. and to allow more variation, specifically the dependence on
We hope that through our work we can create an ap- β, g, and ` variables were not work out of the mathematics.
proachable way for children to understand the complex
mechanics behind the efficiency of these moments. Our
goal is to replicate a Tinkertoy model of the simple walker
originally created by a group at Cornell University [2].

II. THEORY

The first goal of our Clare Boothe Luce work was to

solve for the equations of motion and the jump conditions to define the position of each of the key three spots -
of a passive two-dimensional walker. First, we set up the the two feet masses m and the hip mass M - as shown
angles of the walker and the vectors that we would use in Figure 1. These specifications allow use to define the
 2

constraints of the system as α2 = −φ + π, because the Cornell group we were com-

paring our research to used these angles instead. Also,
`~1 = `{− sin(α1 (t)), cos(α1 (t)), 0} (1) we made this change because the jump condition is best
`~2 = `{− sin(α1 (t) + α2 (t)), cos(α1 (t) + α2 (t)), 0} (2) expressed in terms of θ and φ, as shown in Equation 10.
With the equations of motion we could plot only one
`~M = `~1 (3) stride because in order to have a continuous walking mo-
`~m = `~1 + `~2 , (4) tion there has to be moments where the stance and swing
leg switch and they become the other. This moment is
where ` is the length of each leg and the α angles and where the jump conditions occur in our simulation. The
the vectors are as defined in the Figure 1. approximate moment of this jump is when
From there we solved for the Lagrangian L = T − V
of the constrained double pendulum system, where we
define the kinetic energy T to be φ(t) − 2θ(t) = 0. [4] (10)

This is because the swing leg has reached a stretched

`2 out point where the weight of the body starts to lean on
T = ((2m + M + 2m cos(α2 (t)))α1 0 (t)2 it. The jump conditions that we found were
2
α2 (t)
+ 4m cos2 ( )α1 0 (t)α2 0 (t) + mα2 0 (t)2 )) (5)
2 θ+ (t) → −θ− (t)
 
0
0 2θ − (t) cos(φ− (t))
and the potential energy V to be  θ+ (t) → − −β+β cos(4θ − (t))−2

(11)
 
φ+ (t) → −2θ− (t)
 
 
−0
0 (t) sin2 (θ − (t)) cos(φ− (t))
V = g`((m + M ) cos(α1 (t)) + m cos(α1 (t) + α2 (t))), (6) φ+ (t) → − 4θ −β+β cos(4θ − (t))−2

and got where the pluses denote after the jump conditions have
been applied and the minuses denote the moment before
they have been applied [12]. These jump conditions and
1 equations were then applied to an Objective-C code, as
L= `(−2g((m+M ) cos(α1 (t))+m cos(α1 (t)+α2 (t)))
2 shown in the Appendix.
+ `α1 0 (t)2 (2m cos(α2 (t)) + 2m + M ) In Mariano Garcia’s 1998 Thesis, they expanded the
α2 (t) equations of motion and jump conditions into three di-
+ 4m`α1 0 (t)α2 0 (t) cos2 ( ) + m`α2 0 (t)2 ) [12]. (7) mensions [4]. These mathematics are much more com-
2 plicated and sensitive than those that we investigated
From there we used the Euler-Lagrange equations to in the two-dimensional model. The theoretical model of
solve for the equations of motion. The first equation was a three-dimensional walker is shown in Figure 2. The
most important part about this three-dimensional model
is that in the Cornell group’s paper “An uncontrolled
1 walking toy,” they mention that the simplified theoreti-
θ̈ = (g((β + 1) sin(γ − θ)
`(β cos2 (φ) − β − 1) cal model does not predict the actual motion of the Tin-
− β cos(φ) sin(φ + γ − θ))β`φ̇2 sin(φ) kertoy model [2].

+ 2β`φ̇θ̇ sin(φ) + β`θ̇2 sin(φ)(cos(φ) − 1)), (8)

III. PROCEDURE
where β = m/M and γ is the slope angle. The other
equation was
The setup of our experiment, shown in Figures 3 and 4,
is straightforward and discussed in detail in multiple pa-
1 φ pers written by the robotics group at Cornell [2]. We used
φ̈ = (g(2(β + 1) sin2 ( ) sin(γ − θ) plastic Tinkertoy pieces to create the toy model shown
`(β cos2 (φ) − β − 1) 2
in Figure 4, as opposed to the Cornell group that had
+ (β(− cos(φ)) + β + 1) sin(φ + γ − θ)) wooden pieces available to them, as shown in Figure 3.
φ Additionally, we created a sloped platform for the robot
+ β`φ̇2 sin(φ)(cos(φ) − 1) + 4β`φ̇θ̇ sin2 ( ) sin(φ)
2 to walk down according to the specifications from the
+ `θ̇2 sin(φ)(2β cos(φ) − 2β − 1)) [12]. (9) paper [2]. One difference in our experiment from the
original was that we used a plank of wood and a plas-
Originally, we defined our terms with α1 and α2 be- tic cylinder to prop it up to create our slope. Using the
cause it was the simplest approach for us to figure out plastic cylinder we were able to easily manipulate the
the constraints on the system. We converted these to θ angle if we wished to. After successfully creating this
and φ for our equations of motion, using α1 = θ − γ and toy as close to the original measurements as possible, we
 3

FIG. 2. Theoretical setup of a passive walker. This diagram is

from the Cornell group’s paper “Prediction of stable walking FIG. 4. Diagram of the Tinkertoy passive walking robot cre-
for a toy that cannot stand” and it connects the dimensions ated by the Wooster group. This picture specifies the mea-
discussed in the Cornell group’s theory and computational surements of each part of our passive walking robot. We used
papers to their physical toy model [1]. This group eventually the Cornell group’s diagram as a base for our own and only a
discovered that their simple simulations did not accurately measurements varied [2]. The overall mass of our walker was
predict the behavior of their passive robot. 145.79 ± 0.03 g with the average of the masses on the foot
balances being 41.08 ± 0.02 g.

IV. RESULTS

Creating our passive robot required tinkering with the

set that we had. We were using the specs for a robot
that had been built when wooden Tinkertoys were read-
ily available. In today’s market, wooden Tinkertoys are
collectables and the mass produced toys are plastic. Also,
the pieces that came in our set are slightly different sizes
than the ones in the wooden sets, as you can tell by com-
paring the measurements on Figures 3 and 4. The orange
pieces had to be hollowed out and the spacing for most
of the parts had to be altered, though the overall look is
the same. One of the big differences is the mass distri-
bution. Of the 120 g of the Cornell toy, approximately
100 g was the masses add to the foot balances [2]. In our
toy, though, of the 146 g approximately 82 g was the foot
masses.
FIG. 3. Diagram of the Tinkertoy passive walking robot cre- We then compared our results for the step time and
ated by the Cornell group. This picture specifies the mea- step distance to the original Tinkertoy’s average mea-
surements of each part necessary to create a stable passive surements, as shown in Figures 5 and 6. From these
walking robot. We used this as a start-off point to create
graphs we can tell that the average step time was shorter
our initial robot. [2]. The overall mass of their walker was
approximately 120 g with average foot balances of 50 g.
than that of the wooden Tinkertoy created by the Cor-
nell group. We found that our average step time was
0.34 ± 0.04 s and the Cornell group’s step time of 0.47 s
falls way outside of that uncertainty range [2]. We also
found that our average step distance was 1.8 ± 0.3 cm,
while the Cornell group measured a step distance of 1.3
cm [2]. Though our uncertainty was large in our distance
measured the step sizes and time lengths using a motion measurements, the Cornell step length was still well out-
tracking program. side of our uncertainty range.
 4

0.7 sured the length in time and distance of each step. From
this we were able to compare the behavior of our walker
0.6
to the Cornell group’s wooden walker. The percent dif-
0.5 ference we calculated between our group and the Cornell
)s( emiT petS

0.4
group’s step times and distances was 32% for both val-
ues. We discovered that though the designs were similar,
0.3
the small differences led to noticeable changes in system
0.2
behavior.
There are many reasons that our model toy behaved
0.1
differently from the Cornell group’s toy. One reason is
0.0 that mass distribution due to the differences in materials
0 2 4 6 8 10 12 14
effected the mobility of certain parts of the toy. Also, we
Run Number used less slightly less mass on the feet balances than the
Cornell group (due to material variances), so the bal-
FIG. 5. Time length of steps versus the run number. The ance differences likely caused the robot to walk differ-
red dot on this graph represents the amount of time it took ently. Lastly, we also did not put brass strips on the feet
the Cornell group’s walker to take a step, while the blue dots of our walker and used a wooden slope instead of a metal
represent our data runs, which consist of separate times that one. The inability to stand in any configuration like the
the walker moved down the slope (varying from 1-3 steps per Cornell’s walker may have been integral to the walker
run). The error bars come from our estimated error in mea- stably walking down the hill, but it is unclear.
surements. Our average step time was 0.34 ± 0.04 s, while
the Cornell group’s step time was 0.47 s [2]. This graph al-
Throughout our research, we have learned that from
lows us to visualize the difference between these two walkers’ simulation to physical reality, a passive walking robot
behaviors. is extremely sensitive to initial conditions and param-
eters. Without a specific combination of these things,
the walker lacks stability and will either fall over or get
stuck in a position. We’ve also come to our own real-
3.5
ization that for this kind of system approximations and
3.0 simplifications do not accurately predict or describe the
system. The more complex the model is, the closer it
)mc( ecnatsiD petS

2.5
gets to showing accurate predictions.
2.0

1.5

VI. FUTURE WORK

1.0

0.5
There are many components of this project that we
0.0 were unable to complete because of the time restrictions
of Junior IS. To expand upon the motion analyzed in this
0 2 4 6 8

Run Number
report, we would like to further analyze the motion of
the walker. First, we would expand to looking at the hip
FIG. 6. Distance of each step down the slope versus the run and wobble motion of the walker. To do this, we would
number. The red dot on this graph represents the distance of need to contact the Cornell group to see if they have
the Cornell group’s walker step, while the blue dots represent any motion tracking data or side film from their passive
our data runs, which consist of separate times that the walker walker that we could compare my videos to. Since that
moved down the slope (varying from 1-3 steps per run). The material was not readily available during this research
error bars come from our estimated error in measurements. period we were only able to look at step time and size.
This plot has fewer points than Figure 5 because some of the We would also like to expand this project more into
walker shots were too angled to get a proper measurement.
its education research side. We would acquire multiple
Our average step distance was 1.8 ± 0.3 cm, while the Cornell
group’s step time was 1.3 cm [2]. sets of the Tinkertoy parts that we use for our walker,
we would mark the necessary angles for placements, and
time how long it takes for people of multiple age ranges to
complete building it. From this we should be able to tell
whether this is an efficient base to create an outreach or
V. CONCLUSION
science day workshop out of it. If we can not modify it to
allow for building to be used in outreach, we could create
We created a passive walking robot out of plastic Tin- a fun interactive science day stop where kids can partic-
kertoys, thus creating a physical representation of our ipate in the “Walker Challenge.” This challenge would
computational simple walker model. Then, we filmed the use our current walker, which can only take a few steps,
walker moving down the slope multiple times and mea- and kids would compete to see who could get the walker
 5

to take the most amount of steps down the slope. This VII. ACKNOWLEDGEMENTS
would be a great way to capture kid’s attention during
science day and could be placed somewhere that would
redirect traffic as needed.
This work will be extended into a Senior IS project
as well. We started this research project as a simple
two-dimensional simulation, expanded it into a three- I would like to thank Dr. John F. Lindner, Dr. Cody
dimensional Tinkertoy model, and will take it to a full- Leary, and the Wooster Physics Department for helping
scale dancer moving under the force of various gravities. me with my experiment and providing the supplies nec-
Our hope is that this two year project will overall supply essary for this investigation. Additional thanks to my fel-
the scientific community with a greater understanding of low Junior I.S. classmates and my favorite K-Pop bands
the efficiency of different movements and the connection for keeping my spirit high through the rough patches we
between physics and dance. all experienced the past few weeks.

[1] Michael J. Coleman, Mariano Garcia, Katja Mombaur, [7] Ning Liu, Junfeng Li, and Tianshu Wang. Passive walker
and Andy Ruina. Prediction of stable walking for a toy that can walk down steps: simulations and experiments.
that cannot stand. Physical Review E, 64(2), 2001. Acta Mechanica Sinica, 24(5):569–573, Mar 2008.
[2] Michael J. Coleman and Andy Ruina. An uncontrolled [8] Leslie Philip Pook. Understanding pendulums: a brief
walking toy that cannot stand still. Physical Review Let- introduction. Springer Verlag, 2013.
ters, 80(16):3658–3661, 1998. [9] Maximo Roa, Diego Garzon, and Ricardo Ramirez.
[3] S. H. Collins. A three-dimensional passive-dynamic walk- Climbing and Walking Robots: towards New Applica-
ing robot with two legs and knees. The International tions. InTech, 2007.
Journal of Robotics Research, 20(7):607–615, Jan 2001. [10] Amazon Toys. Tinkertoy 30 model super building set –
[4] Mariano Garcia. Mariano Garcia’s PhD Thesis. PhD 200 pieces – for ages 3 preschool educational toy.
thesis, Cornell University, 1998. [11] M. Vukobratovic. Robot-environment dynamic interac-
[5] Mariano Garcia. The simplest walking model: Stabil- tion survey and future trends. Journal of Computer and
ity, complexity, and scaling. Journal of Biomechanical Systems Sciences International, 49(2):329–342, 2010.
Engineering, 120(2):281, Jan 1998. [12] Justine Walker and John F. Lindner. Clare boothe luce
[6] E. Atlee Jackson. Exploring nature’s dynamics. Wiley, research mathematica notebook. Contains all of the
2001. derivations of the equations of motion and the jump con-
ditions, February 2017.