You are on page 1of 3

A toy climbing robot

Matthew Bell Devin Balkcom


Dartmouth Computer Science Department Dartmouth Computer Science Department
Hanover, NH 03755 Hanover, NH 03755
mbell@cs.dartmouth.edu devin@cs.dartmouth.edu

Abstract— We built a simple toy climbing robot in order to


explore problems related to minimalist grasping, path planning,
and robot control. The robot is capable of climbing a wall of
pegs either under remote control, or on the basis of a set of
pre-recorded keyframes. In addition, the robot can climb certain
peg configurations using a cyclic gait. All communications are
sent through an infrared connection, and the tether to the robot
consists only of two power wires. Due to the minimalist, non-
prehensile grasping method, the robot is capable of actively
removing error while climbing, which is necessary to enable the
robot to climb without sensing the environment.

I. I NTRODUCTION
Our goal is to develop a simple, lightweight robot that
uses minimal computation and sensing in order to successfully
climb a wall of pegs. Our robot utilizes minimalist grasping
Fig. 1. The robot on the climbing wall
and a non-prehensile grip, allowing it to slide on the pegs to
actively remove error. Most robot grasping problems involve
grasping a small object with a large industrial manipulator; our
cone is empty, meaning that no free motions are possible and
robot can be thought of as a small manipulator grasping the
the robot is stable.
entire climbing wall. Due to the lack of environmental sensing,
Our toy robot is largely a preliminary exploration into
the robot cannot tell if it is off course; it is thus necessary to
the challenges and limitations involved in building a simple
plan motions in a way that provides stability and repeatability.
climbing robot. We explored several problems related to path
A major goal was to keep the robot as simple as possible to
planning and minimalist grasping under uncertain conditions,
make it feasible for the general public to buy an inexpensive
especially in the area of error removal.
kit for building the robot. Our robot is made of hobby servos
and LEGO pieces (See Figure 1). There are three climbing II. R ELATED W ORK
modes:
Our toy robot is not the first climbing robot; it is, however,
1) Manual remote control the simplest, one of the lightest, and the only one to reliably
2) Autonomous, with pre-recorded keyframes remove error while climbing open-loop without sensing the
3) Autonomous, using a simple cyclic gait environment. Rus and Kotay developed the Inchworm [4],
The robot is capable of climbing under manual control through a lightweight task-reconfigurable robot capable of climbing
a Java interface and autonomously using a set of pre-recorded on any ferrous surface that it can grasp with electromagnets.
keyframe positions. For appropriate wall configurations, a set Michigan State University has a pair of climbing robots
of cyclic keyframes exists that will make the robot climb that use suction to climb on smooth walls and ceilings [9].
the wall with a cyclic gait. When executing pre-recorded Nagakubo and Hirose [5] built a large, heavy quadruped robot
keyframes, the robot climbs open-loop, with no sensor feed- capable of navigation on horizontal and vertical surfaces using
back. suction. Neubauer [6] developed a small robot for climbing
Our mathematical model of the robot considers the arms and inside pipes. Neubauer’s robot is similar to ours in its use of
legs to be rigid, and assumes that the servos can be locked into a non-prehensile grip. Linder and a group of undergraduate
position. We derived the forward and inverse kinematics for students developed Tenzing [7], a large quadruped robot.
use in the GUI, and used Reuleaux’s method [1] to analyze Tenzing is heavier than the toy robot and employs significant
stability geometrically to gain an intuitive understanding of sensing, including live video, to assist in motion planning.
stable grasps. We represent the free motions of the robot with Bretl at Stanford has developed a path-planning algorithm for
a polyhedral convex cone, represented in matrix form. Using JPL’s LEMUR II robot [8]. This robot is capable of climbing
Goldman [2] and Hirai [3], it is possible to determine if this walls with arbitrarily shaped and angled handholds. The design
D
C
Θc2 Θd2
pd2
l2
l1
Θc1
Θd1
γ r
δ (a) Initial position (b) Legs lifted
α
(x,y,Θ)

A B

Fig. 2. Variables used in mathematical analysis

(c) Recentering position (d) Arms lifted

of this robot is such that there is no definite “up” orientation Fig. 3. Sequence of robot configurations during a single cycle of a gait
of the robot body, providing more flexible manueverability.
III. ROBOT D ESIGN
contact normals are given by a matrix N , then the free motions
The robot is built from eight small hobby servos and several
of the robot are given by the set {q̇ : JN q̇ ≥ 0}. If this set
LEGO pieces, and is controlled by a Pontech SV203C board.
is empty, then the robot is motionless and stable. This set
The robot receives all of its control communications through
represents a polyhedral convex cone, and we can determine if
an infrared (IR) receiver, and as a result, the robot’s tether the cone is empty.
consists only of two wires for power.
The IR communications rely on three main pieces of V. C LIMBING M ODES
software. The host system runs a Java application that sends
The robot is capable of climbing either under manual remote
commands through a serial cable to the standalone transmitter
control, or using interpolating motions between pairs of a
SV203C (Tx board). Embedded code runs on the Tx board,
sequence of pre-recorded keyframes. In addition, a cyclic
and retransmits the commands along with a checksum to the
gait has been developed for autonomously climbing a vertical
receiving SV203C (Rx board, located on the robot). Additional
ladder configuration of the wall (See Figure 3).
embedded code on the Rx board interprets and verifies the
commands before moving the appropriate servo. A. Pre-recorded Keyframes
IV. M ATHEMATICAL A NALYSIS The robot’s primary climbing method uses a series of pre-
recorded keyframes to guide the robot up the wall. These
Using the variable definitions given in Figure 2, we analyzed
keyframes are recorded by the human operator while the
the forward and inverse kinematics of the robot. Each arm has
robot is being navigated up the wall under manual remote
2 degrees of freedom, which are indicated by θx1 and θx2 ,
control. During playback, interpolation between the keyframes
where x is the arm label. The position of arm D, represented
is handled by moving all the arms at a constant angular rate
by pd2 , is given by
until they reach the position specified by the next keyframe,
resulting in a much smoother and quicker climb than under
 
x + rcδ + l1 cd1 + l2 cd1 d2
pd2 = Rθ (1)
y + rsδ + l1 sd1 + l2 sd1 d2 manual control. The actual speed is dependent on the route
chosen by the human operator.
where Rθ is a rotation matrix, cij... = cos θ1 + θ2 + . . ., and
It was generally necessary to position the robot within about
sij... = sin θ1 + θ2 + . . ..
0.5 cm horizontally of the initial position used when recording
The inverse kinematics for one arm are computed according
the keyframes to ensure consistent success. Small variations
to standard methods. However, we only consider angles in the
in position are insignificant due to the robot’s non-prehensile
ranges 0 ≤ θ1 ≤ π and 0 ≤ θ2 ≤ π to reduce the number of
grip, as the robot will not fall as long as some portion of
solutions, as we have observed that these constraints hold for
each of the four limbs is touching a peg. Introducing error-
climbing methods with the most stability.
correcting motions into the keyframes can increase the amount
of acceptable variation of the initial position.
 
l22 −l21 −x2 −y 2
θ1 = arccos √ + arctan2 (y, x)
−2l1 x2 +y 2
 2 2 2 2 (2) B. Cyclic Gait
x +y −l1 −l2
θ2 = arccos
If the pegs are in a repeating configuration, it is possible to
2l1 l2

To compute the free motions of the robot, we first computed climb using a cyclic gait. The configuration that we specifically
the Jacobian matrix J from the forward kinematics. If the examined is a ladder formed from two sets of pegs in vertical
be possible for the robot to wave its arms around to determine
where handholds are in its immediate vicinity, and to use this
information to climb locally. From this local climbing ability,
it should be possible to recursively climb the entire wall. The
robot can climb locally until it is unable to do so, and then it
can back down the wall for some distance, and try climbing in
a different direction. As it proceeds, it will slowly develop a
model of the entire wall from the local information it collects.
Fig. 4. Position used to recenter the robot during a gait cycle
VII. C ONCLUSIONS
A toy climbing robot was successfully developed. Although
lines. The robot can climb this configuration at an upward it is not capable of automatically pre-planned climbing, it can
speed of about 4.1 cm/min. This type of configuration greatly climb a ladder using a cyclic gait in an open-loop mode. The
simplifies the problem of finding a path up the wall, as it robot achieves this through the use of a recentering motion
is only necessary to find one complete cycle. For a ladder that ensures that the robot is correctly positioned during each
with perfectly vertical sides, the cycle only needs to result in cycle. The robot is sufficiently simple that it could be marketed
each arm of the robot being one handhold higher up the wall as a kit—it does not require a bulky tether to operate correctly,
(Figure 3). We developed a gait for the ladder configuration and the host software’s GUI is fairly easy to use.
by manually controlling the robot through one full cycle, This project has also explored the mathematics involved in
and recording the keyframes from this cycle. The motions robotic climbing, including the forward and inverse kinemat-
designated by the keyframes were then repeated multiple times ics, and calculations of stability. The cyclic gait method has
to drive the robot up through several cycles of the gait. been successfully implemented for one wall configuration, and
In order for a gait to be successful, the robot must not other gaits could be developed for other regular configurations.
shift from side to side as it climbs up the wall. The fact It is possible that a general gait could be developed for a set
that the robot is climbing open-loop in this mode requires of similar configurations, provided that a generic recentering
robust trajectories to ensure success. Thus, it is necessary to strategy exists.
perform some motion or sequence of motions that returns the There are several potential areas of further exploration. A
robot to some known configuration. The frequency of this full path planning algorithm could be developed to allow the
error-correcting motion depends on the complexity of the wall robot to climb any wall based only on knowledge of the
configuration; however, it should occur at least once in every locations of the pegs. A general-purpose cyclic gait could exist
cycle. for a certain set of wall configurations. Sensors could be added
For the ladder configuration that we examined, we devel- and incorporated into the climbing algorithm, and the robot
oped a maneuver that will successfully recenter the robot itself could be redesigned to allow it to climb blindly without
into the position shown in Figure 4. In order to achieve this any knowledge of the wall.
position, the robot hooks its upper arms over the handholds
at angles that cause the robot to slide until the handholds R EFERENCES
are at the elbow joints. For other configurations, it may be [1] F. Reuleaux, The Kinematics of Machinery. MacMillan, 1876, reprinted
necessary for the robot to fall slightly in order to become by Dover, 1963.
[2] A. J. Goldman and A. W. Tucker, “Polyhedral convex cones,” in Linear
recentered. These recentering methods require the robot to take Inequalities and Related Systems, H. W. Kuhn and A. W. Tucker, Eds.
some action in order to remove error from the system. Due to York: Princeton Univ., 1956, pp. 19–40.
the recentering manuever executed during the gait, the robot [3] S. Hirai, “Analysis and planning of manipulation using the theory of
polyhedral convex cones,” Ph.D. dissertation, Kyoto University, Mar.
is much less sensitive to the intial position when climbing in 1991.
this mode. [4] K. Kotay and D. Rus, “The inchworm robot: A multi-functional system.”
Auton. Robots, vol. 8, no. 1, pp. 53–69, 2000.
[5] A. Nagakubo and S. Hirose, “Walking and running of the quadruped
VI. O PEN P ROBLEMS wall-climbing robot,” in IEEE International Conference on Robotics and
Automation, vol. 2, 1994, pp. 1005–1012.
A major source of possible failure for the cyclic gait is [6] W. Neubauer, “A spider-like robot that climbs vertically in ducts or
an improper recentering motion. Adding a minimal sensor pipes,” in IEEE/RSJ/GI International Conference on Intelligent Robots
package to the robot would permit detection of the correctness and Systems, vol. 2, 1994, pp. 1178–1185.
[7] S. P. Linder, E. Wei, and A. Clay, “Robotic rock climbing using computer
of the recentering motion. However, another solution is to vision and force feedback,” in IEEE International Conference on Robotics
make use of gravity. In Figure 4, if the angles of the arms and Automation, 2005.
with respect to the horizontal are steep enough, the robot will [8] T. Bretl, S. Rock, J. C. Latombe, B. Kennedy, and H. Aghazarian, “Free-
climbing with a multi-use robot,” in International Symposium on Robotics
overcome static friction and slide into the recentered position. Research, 2004.
An interesting area for further research is in climbing [9] J. Xiao, J. Xiao, and N. Xi, “Minimal power control of a miniature
using only local information. Full path planning requires prior climbing robot,” in IEEE/ASME International Conference on Advanced
Intelligence Mechatronics, July 2003, pp. 616–621.
knowledge of the entire wall, which may not be available. With
either touch sensors or servo torque measurements, it should

You might also like