IEEE Robotics & Automation Magazine - December 2015 - 21

For human-sized robots, a pioneering work, and the current state of the art, has been done in [17] for a traditional
Japanese dance based on IK and the LFO. It allows the interaction with a motion designer who can set up some key-frame
postures that stabilize and guide the numerical optimization
[16]. The method leads to impressive dynamic movements,
coming from its noncausal methodology. The downside is that
it is difficult to do in real time, for example, when dynamic
walking is needed. This noncausality is needed to capture a
prediction horizon, which makes the achievement of particular dynamic movements possible, as emphasized by the study
of walking [1]. A new step must be anticipated, implying an
action on the past trajectory, when looking at the motion as a
trajectory optimization problem.
Since OSID is an instantaneous-linearization method, it
presents similar limitations, but it can be applied in real
time without interactions with an expert during the movement adaptation for most movements. Moreover, different
constraints, such as joint limits and collision avoidance, can
be considered, which is a significant advantage over the existing methods. As explained in the introduction, this application shows that OSID is an appealing tradeoff between
IK and trajectory optimization, providing both the control
of a large range of movements and fast computations.
However, as described in the "Balance of the Robot" section, the method is not able to independently generate
movements that require a prediction [27], such as a walking
step, due to the instantaneous linearization. Similar to IK,
this type of effect should be specifically integrated [18],
which implies that the trigger of a new step must be anticipated by a prediction filter [1] if OSID-based imitation
should be implemented in real time.

nearly pure white noise and negligible bias. More than thirty
minutes of movements were captured for the preparation of
the show. The captured data for the most important motions
in this work is freely available (http://projects.laas.fr/gepetto/
novela/noveladb) along with the geometric retargeting and
the results of the dynamic SoT, explained in the "Dynamical
Retargeting" section.
Joint trajectories obtained by geometrical retargeting are
not dynamically consistent with the robot model, since nothing guarantees that the robot is stably balanced or that autocollision or joint limits are avoided. Moreover, some important
aspects of the original motion can be damaged, since the retargeting is obtained by a tradeoff among the positions of all the
bodies. If a given body is more relevant than the others, this
importance is not reflected in the obtained motion due to the
differences between the two kinematic chains. For example, if
both hands are clamping in the demonstration, their resulting
positions are not likely to satisfy the clamp.
Dynamical Retargeting
The SoT is a means to enforce dynamic consistency while simultaneously editing specific aspects of the resulting motion.
It is possible to make the robot more precisely track parts of
the demonstration by adding a task on specific operational
points to follow the exact corresponding trajectory of the
human performer. This section describes the generic retargeting and editing process, shown in Figure 3.
The contact points are extracted by detecting clusters of
static points in the feet trajectories from the demonstration.
The first constraint of the SoT is to enforce dynamic consistency of the corresponding contact model. This also guarantees local balance so that all contact points remain stable.
Other robot constraints are added to enforce joint limits or
collision avoidance. The lowest priority task, the posture task,
tracks the reference configuration coming from the geometrical retargeting in the best possible way. This task ensures that
the global structure of the movement looks similar to the one
that has been demonstrated. However, since the kinematic
structures of the demonstrator and the robot are different, the
posture similarity (corresponding to a least-norm tradeoff) is
sometimes not satisfactory. Typically, some important aspects

Data Acquisition
The motion capture system provides the spatial trajectory for
each of the markers attached to the human body, as shown in
Figure 2. The markers are distributed on the human body in
positions to minimize the motion of the skin with respect to
the bones and to facilitate temporal tracking. These markers
are manually associated to form a skeleton in which every
link (or bone) is characterized by its position and orientation
in a fixed frame. The robot configuration is
then adjusted to achieve the best fit to the
Simulation/
measured body positions and orientations.
Motion
Posture Task
Robot
Capture
The geometry is retargeted by solving a nonSoT
linear least-squares problem, minimizing
Pattern
the distance between the observations and
Footsteps
Generator
the model for each time frame [22].
The human motion was acquired using a
Operational
Motion Analysis motion capture system
Point Task
(http://www.motionanalysis.com/) made up
Manual
of ten infrared cameras distributed around
Choreography
the experimentation zone and calibrated by
Motion Analysis software (http://www.moFigure 3. A dynamical retargeting scheme. Motion capture information passes to the SoT
tionanalysis.com/) with an acquisition fre- through various tasks. Tasks for operational points can be manually activated according to
quency of 200 Hz and precision of 2 mm of the choreography. The motion is finally executed in a dynamic simulator or the real robot.
DECEMBER 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

http://projects.laas.fr/gepetto/ http://www.motionanalysis.com/ http://www.mo http://www.tionanalysis.com/

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2015