(Kujawski and Bower, 1993) and studies of 3 and 5
year olds have shown that both can perform reliable
recognition of human and nonhuman forms at 3 years
of age with the 5-year-olds exhibiting ceiling levels of
recognition. Moreover, there is evidence to suggest
that 2-year-olds can utilize subtle distinctions in the
kinematics of movement styles to determine the dif-
ference between real and pretense movements (Lil-
lard and Witherington). Thus, we can consider the
abilities of adults to exhibit a fine-tuning and elab-
oration of the extremely sophisticated mechanisms
already evident in the youngest infants.
Besides its development, there are various other
aspects of human movement perception which sug-
gest that human movement is a special class
of motion with specialized mechanisms devoted
to its processing. One question which has ad-
dressed the issue of what is special about hu-
man movement has been whether the critical
mechanism for the perception of human move-
ment relies on specialized low-level motion detec-
tors (Mather et al., 1992, Neri et al., 1998) or on
higher-level mechanisms which organize the results of
low-level motion detectors (Shiffrar and Freyd, 1990,
Thornton et al., 1998). The resolution of this ques-
tion has relevance for the development of motor pro-
duction techniques for humanoid robots. For exam-
ple, any systematic change between when a move-
ment is evaluated against low-level physical prop-
erties and when a movement is evaluated against
more cognitive aspects would be relevant. One ex-
ample of such a systematic change in perception is
demonstrated by the perception of two-frame appar-
ent motion sequences of human movement. If the
timing between frames is short then observers re-
port seeing impossible movements where limbs ap-
pear to interpenetrate each other during a move-
ment. As the timing between frames increases a
solidity constraint is imposed and the movement is
perceived as two limbs moving around each other
(Shiffrar and Freyd, 1990). It is hoped that study-
ing the perception of humanoid movement might re-
veal unique aspects of human visual perception and
visual cognition.
3. Production of humanoid move-
ments
In this section we introduce the relevant theories of
motor production, discuss their implementation and
then present how stimuli were obtained for use in the
perceptual experiment.
3.1 Theoriesofmotorproduction
There are a large number of theories regarding
the way in which the brain plans and orchestrates
physical movements of the body. Theories vary
according to the space in which planning occurs
(intrinsic body coordinates or extrinsic environment
coordinates), whether planning is purely kinematic
or dynamic, and to what the level the theory of
motor production incorporates the physical inter-
action of the body with its environment through
biomechanical and physiological principles. A
consideration of extrinsic kinematic trajectory
planning lead to the minimum jerk hypothesis
(Flash, 1983, Hogan, 1984, Flash and Hogan, 1985)
according to which motions are planned as tra-
jectories in the physical environment satisfying
minimization of the third derivative. Snap mini-
mization was also investigated (Flash, 1983) but
a general property of extrinsic kinematic planning
is that point-to-point motions yield straight paths
inconsistent with some empirical findings which have
detected a curvature in point-point reaching motions
(Atkeson and Hollerbach, 1985, Uno et al., 1989a,
Haggard and Richardson, 1996). The minimum
kinaesthetic jerk model (Wann et al., 1988) circum-
vents this complaint by modeling the instantaneous
stretch and limb’s centers of mass yielding dynam-
ically planned motion with minimal internal and
external jerk. Alternatively, the minimum jerk
virtual trajectory model (Flash, 1987) proposes that
motion is planned using a minimum jerk trajectory
but that this virtual trajectory acts as a guide for the
actual limbs in the form of a spring-damper system
between actual and target joint angles. Other re-
searchers have attempted to explain the curvature of
hand trajectories, proposing that motion is planned
in intrinsic kinematic space according to straight
paths in joint-angle space (Soechting et al., 1986), or
that motion is planned as a straight path in percep-
tually distorted visual space (Wolpert et al., 1994).
While animators have used the intrinsic dynamic
method of minimum torque (Rose et al., 1996), it
has been proposed that motions are indeed planned
according to the minimization of torque change
throughout motion (Uno et al., 1989a). Higher level
models have extended this concept by proposing the
motion is planned according to the minimization
of muscle tension change (Uno et al., 1989b) or
neural activation change. It has also been proposed
that planning occurs as an optimization at the
control signal level with a signal noise component
proportional to the activation level and that the
constraint is minimum end-point positioning error
(Harris and Wolpert, 1998). Finally, a class of
theories based on the equilibrium point hypothesis
(Feldman, 1966) suppose that the body is controlled
using equilibrium positions at which the forces gen-
erated by the muscles supporting a given joint are
in balance while the actual muscle force results from
a lower level of control based on these equilibrium
parameters (Bizzi et al., 1976, Abend et al., 1982,