Axa Assurance Maroc - Insurer Innovation Award 2024
Poster iros2019
1. Universidad Carlos III de Madrid
Whole-Body Postural Control Approach based on
Multimple ZMP Evaluation in Humanoid Robots
J. M. Garcia-Haro, S. Martinez, E. D. Oña, J. G. Victores, and C. Balaguer
4. WBPC ARCHITECTURE
5. CONCLUSIONS
1. OBJECTIVE
This article presents a postural control approach of a
humanoid waiter robot. In previous research, two methods
were proposed to address this complexity. The first one was
an improvement for the control of body balance
(locomotion). The second one was a method to apply the
classical concepts of body balance to transport objects in a
tray (manipulation).
This approach is based on the concept of a multi-ZMP
evaluation system to control the object and the body
stabilities. Both methods were developed independently,
avoiding disturbances between them. The integration into a
whole-body postural control architecture is a challenge for the
performance of both methods, due to the significant influence
between them. In this research, we present both methods to
deal with the complexity of the humanoid task.
2. BODY BALANCE CONTROL
There are mechatronic problems
and also related to the body balance
task that can be notable during the
stability control. These errors arise
from the linearization of this robot
model itself, the approximation of
CoM, measurement deviations in
the sensors, the structure flexibility,
irregularities in the ground.
Based on the open-loop set Push-Recovery experiments, F-T sensor measurements are
captured and processed to model the ZMP error. The idea is to modify the LIPM model. The
new model called DLIPM (Dynamical Linear Inverted Pendulum Model) adds
dynamically the modelled error in every control loop, compensating both the static error (ka)
and the transitory dynamic state limiting oscillations (Ba) [1].
3. OBJECT BALANCE CONTROL
A non-grasping manipulation task is a common
option for transporting objects and has certain
advantages. But the problem for this object
balance control is the tray orientation, and
therefore, of the measurements of the F-T
sensor. In order to apply the ZMP equations, the
tray must always be coplanar with the
horizontal plane. Only in this way, all the forces
and torques exerted by the bottle will be
correctly reflected in the sensor. In the case of
the TEO robot, the tray will have different
orientations during the balance control. These
poses, related to the state of the bottle, generate
data on the sensor that cannot be applied
directly to the ZMP computation. For the use
of the ZMP equations, it is necessary to apply
transformations based on the pose estimation
concept of 3D Dynamic Slopes [2].
The approach for a waiter
robot aims to transport
objects (drinks or food) on
a tray. Therefore, this robot
must have the following
primary skills. On the one
hand, the robot must
maintain its balance
(walking tasks). On the
other hand, the robot must
transport objects on a tray
(handling tasks without
grasp).
The WBPC architecture introduces the evaluation of multiple ZMP (object and body).
Therefore, this approach is based on executing the upper-body and the lower-body controllers
in parallel. In this case, the physical influences are contemplated with systems type FIS
(Fuzzy Inference System) [3]. The reason for use FIS systems is bio-inspiration. We think that
it is interesting to use a system similar to the way human beings think for a human application.
To verify the viability of the proposed system, the WBPC controller is compared using two
models for the body. First with the LIPM model and then with the improved DLIPM model.
The last figure shows the stability control performance for the object (ZMPobj) and the body
(ZMPbody). In these tests, the robot is pushed to check its response. In the first case of the
WBPC using the LIPM model, both controllers over-oscillate and also have a long
stabilization time. The model totally determines the dynamics of equilibrium, and it is not
possible to modify it. In the case of the DLIPM model, the answer is better. ZMPbody
overshoots are smaller and less intense. The time response is four times less. Also, the ZMPobj
stabilizes faster. The reason is that the DLIPM is capable of absorbing both external and
internal disturbances. In this way, the disturbances caused by the arm are minimized, and the
object controller works better.
6. REFERENCES
For the integration of both methods, a WBPC architecture has been developed. This one is
based on the concept of multiple ZMP evaluation and the interrelation of the object and robot
stabilities. Thus, the exposed methods for computational stability consider the influences
between them. It is achieved by developing a diffuse system that simplifies the complexity of
the task. A more in-depth investigation is needed to evaluate the balance during the march, but
the first stage of research in this field has been successfully achieved, as can be seen in the
link: https://youtu.be/XJzGwDIjFCY
[1] Martinez, S., Garcia-Haro, J. M., Victores, J., Jardon, A., & Balaguer, C. (2018).
Experimental Robot Model Adjustments Based on Force-Torque Sensor Information. Sensors,
18(3), 836.
[2] Garcia-Haro, J. M., Martinez, S., & Balaguer, C. (2018). Balance Computation of Objects
Transported on a Tray by a Humanoid Robot Based on 3D Dynamic Slopes. In 2018 IEEE-
RAS 18th International Conference on Humanoid Robots (Humanoids) (pp. 704–709). IEEE.
[3] Hernandez-Vicen, J., Martinez, S., Garcia-Haro, J. M., & Balaguer, C. (2018). Correction
of Visual Perception Based on Neuro-Fuzzy Learning for the Humanoid Robot TEO. Sensors,
18(4), 972.