Professional Documents
Culture Documents
Hubbard 1980 Human Control of The Skateboard
Hubbard 1980 Human Control of The Skateboard
MOW HKRBARD
Department of Mechanical Engineering. University of California, Davis, CA 95616, U.S.A
Abstract Human control of the skateboard is investigated by modeling the rider as a single. rigid bed)
pinned 10 rhe board along the roll axis. Human input is taken to be a torque applied at the ankles. The
equations of the nonholonomic rider-board system are presented and a simple tracking task is established
11s dynamics are augmented by those of the skateboard and rider to describe the complete system in the
tracking mode Several control schemes are discussed. Under certain conditions, simple proportional
feedbach conrrol of rider tilt angle can stabilize roll motion of the vehicle, but in the most general case full
stale feedbach 1s required. In the complete state feedback case, a performance index for the tracking task IS
defined and the minimizing feedback gains determined. Time simulations of the tracking task using the
optimal feedback gains are shown. Experimental results are presented which tend to validate the theory.
745
746 MOhT HUBBARD
and
the full state feedback approach to control law for-
mulation taken in what follows. Although not re- mlu’+ack; T
quired, the value of a is henceforth taken for simplicity
to be zero, and the rider is assumed to be positioned
G=
F O-
(ml* + I,)oc
u/at 0
I
at point P.
HUMAN CONTROL
lated from
For purposes of illustration. a particular rider-
C= -GTSfb (15) skateboard system was analyzed. Shown in Table I are
the inertia and geometric parameters of [he vehicle and
and where S is the only positive definite solution to the
matrix Riccati equation
Table I. Skateboard and ruder parameter\ for numertcal
SF + FTS + A - SGB- ‘G’S = 0. (16) stud!
~_____.
It is often reasonable to choose A diagonal with m 77 bg a 0.57 m
elements aii = l/x$ where xai is the maximum I 0.925 m C 0 75
1, 12.2 kgm’ h 50 N m rdd
allowable value ofthe ith element of the state vector. In
a 0.0 I‘ 2 0 m vx
this case (13) becomes -__-..
750
Moryr l-iUBBARD
rider, which were chosen to be near those of an actual From the preceeding it is clear that the control task
test subject. can be subdivided into two portions; tracking (0 and _r)
Equation (16) was solved by the method ofeigenvec- and balancing (4 and 4). Ahhough neither is neglected
tor decomposition using a computer program adapted entirely by a controller designed using linear-
from the one described in Bryson and Hall (1971). quadratic methods, it is of interest to see the variation
Table 2 gives six different sets of state and control in the optimal closed-loop system as the emphasis
weightings and the corresponding feedback gains and changes from one to the other. Specifically, the perfor-
closed loop system eigenvalues. The effect is shown of mance index can be written as
decreasing the allowable position deviation X,. All the
feedback gains gradually become larger with the
largest increase in Cx (from 0.50 to 3.16). In addition
the dominant (smaller) closed loop system eigenvalue
J=
Iom{(#/dd’ -I- (&d,.,’+ cCCW~.J’
Table 2. Optimal feedback gains and closed-loop eigenvalues for various weight&
1 0.1 0.32 0.1 0.2 0.1 1.32 2.45 1.89 0.50 -5.59. -3.14, -2.26, -1.06
2 0.1 0.32 0.1 0.1 0.1 8.97 3.00 2.13 1.00 -5.37, -3.16, -2.13+0.63j
3 0.1 0.32 0.1 0.07 0.1 10.29 3.44 3.41 1.41 -5.04, -3.18, -2.56kO.91,
4 0.1 0.32 0.1 0.06 0.1 11.27 3.11 3.93 1.73 -4.56, -3.22, -2.97kO.991
5 0.1 0.32 0.1 0.05 0.1 12.09 4.05 4.36 2.00 -3.41?0.17/, -3.62+1.1Oj
6 0.1 0.32 0.1 0.03 0.1 15.55 5.20 6.20 3.16 -3.04+0.15j, -4.62k2.66,
-___ ____-
Human control of the skateboard 751
i -1
Fig. 7. Locus ofclosed-loop system roots as a function of changes in the relative weight Eof the tracking and
balancing subtasks.
of 6-12 rad set-’ (1 to 2 Hz). Thus care must be taken two complex conjugate pairs, with dominant time
in the interpretation of the optimal human controllers constant of roughly 0.5 sec. Hence the decay of the
generated in this fashion. system states to zero is relatively fast, the position error
To serve as an illustration of the actual time decaying to near zero in about 1Ssec.
response of the closed-loop controllers discussed pre- Figure 9 shows the time response ofanother optimal
viously, time simulations were run which generated closed-loop system (with eigenvalues - 1.56 t_ .i 1.17,
solutions to (18) for a variety of initial conditions and - 6.00, - 1.82s~~~~’ generated from weights 4, =
for three different control laws generated from different 0.1 rad, 4. = 0.1 rad/sec, 0. = 0.1 rad, X, = 0.1 m, y,
weighting matrices in (13). These results are shown in = 0.05 rad) to an initial position error X,, = 0.1 m.
Figs. 8-10. In each case $, X and y are plotted vs time. Here two of the closed-loop eigenvalues are real, the
The two remaining state variables can be inferred in dominant time constant is roughly 0.65 set and the
each plot by graphical differentiation. decay of the system states to zero is somewhat slower
Figure 8 shows the closed loop time response for than in Fig. 8. The initial control action is to turn to the
case 6 in Table 2, for an initial rider tilt error do = right (y>O), away from the reference line. Although
0.02 rad. Note that initially the board tilt angle y is this, at first glance, may seem nonoptimal, the con-
positive, producing a right turn to move the board troller is merely unbalancing the rider to the left (4
under the rider. This produces both heading (8) and < 0) so that it can then begin a left turn and approach
position (X) deviations which must then be corrected the reference line. From the positivity of all feedback
for. Note that the closed-loop eigenvalues consist of gains in Table 2, this same behavior can be seen to be
true for isolated Qinitial conditions. The initial control
action is to turn even further away from the line,
I ozc n I increasing 0 momentarily before the eventual cor-
rective action can be taken.
Such a time simulation ISshown in Fig. IO, for case 1
in Table 2 with an initial error 8, = 0.1 rad. Here all
four eigenvalues are real and the dominant time
constant is roughly 1.0 sec. Thus the system response
is slower still.
ESPERIMESTAL RESULTS
p ;:
- 0 --0
-8 #
Fig. 9. Slower closed-loop system response to initial transverse position error with feedback gams C = [6.20,
2.13, 1.53, O.SO].
measured from the film. Position deviation X was had a measured forward velocity, u = 1.79 m/set. The
measured from a ground trace left on the concrete floor closed-loop rider-board system appears to have a
by chalk attached under the board center. speed of response somewhere between those shown
The test subject was an experienced professional in Figs. 9 and 10. Furthermore, the qualitative re-
skateboard rider, the reigning world downhill cham- semblance of the experimentally measured X. C#J and 7
pion at the time. The subject was instructed to actuate to those of the theoretical turn shown in Fig. JO is very
with the ankles only, maintaining the rest of the body good. Initially rider tilt $I is negative but it returns to
as rigid as possible, and to track the final reference line zero at t z 1.3 set and then overshoots slightly. The
as closely as possible. No detailed quantitative specifi- board tilt y, after the brief initial positive impulse
cation of the tracking task was made however, as for discussed previously, follows a similar path becoming
example including desired weighting factors, etc. Rider negative and then returning to zero and overshooting
and board parameters were as listed in Table 1 except slightly. The position error X is always positive
for the speed which varied slightly from test to test. reaching a peak at t 5 0.4sec and then decaying to
Although the experimental study included many near zero after roughly two seconds.
factors, only a nominal set of results are presented here. It should be possible, using sophisticated system
An experimental trajectory is shown in Fig. 11 which identification techniques, to determine the exact feed-
Fig. 10. Slow closed-loop system response to initial heading error (equivalently. system dynamic response
for turning through small angle) with feedback gains C = [7.32, 2.45. 1.89. 0 503
Human control of the skateboard 753
a wheelbase
A state weighting matrix in performance Index
REFERENCES b control weighting parameter In performance
index
Atlenza, F. and Sia, C. (1976) The Hazards of Skateboard- B.E endpoints of desired trajectory
Ridmg, Pediarrics 59, 939-942. constant relating board till angle to steering
Baron, S., Kleinman, D. L. and Levison, W. H. (1970) An angle
Optimal Control Model of Human Response, Part II: C feedback gains
Prediction of Human Performance in a Complex Task, d distance of rider c.m. behind front axle
Auromarica 6, 371-383. F system dynamics matrix
Bellman, R. and Kalaba, R. (1965) Quasilinearizarion and G control distribution matrix
Non-Linear Boundary-Value Problems, Elsevier, New, h board cm height above axles
York mass moment of inertia
754 MONT HUBBARD