In this paper we introduce the SinfonIA Pepper Team for RoboCup Social Standard Platform
Our team is a collaborative effort between Colombian institutions whose main objective is to
the development of robotics and artificial intelligence in our country.
Our current research interest is to improve the human robot interaction by
a) customer service applications based on reliable face detection and person identification
b) entertainment services offering two-player games, based on reinforcement learning.
In this paper, we also show our results implementing a custom version of the Stage 2
tests where is this and stickler for the rules. We use our Pepper robot to demonstrate
abilities such as human-robot interaction, person recognition, localization, navigation and
In this paper we introduce the SinfonIA Pepper Team for
RoboCup Social Standard Platform League (SSPL). Our team is a joint
effort between several Colombian institutions whose main objective is to
lead the development of robotics and articial intelligence in our country,
so that new applications and technologies are introduced in people's daily
life. Our current research interest is to improve the interaction between
robots and humans by a) allowing the robots to learn new instructions
in a natural way using multimodal inputs and b) developing an emotion
recognition system that the robots can use to drive its interaction with
the humans. In this paper, we also show our efforts implementing a shortened
version of the Cocktail Party challenge from previous years, using a
Pepper robot that demonstrates speech interaction, person recognition,
object detection and task planning capabilities.
Qualification Video SinfonÍA Pepper Team @Home (SSPL) RoboCup 2019
Identification of Multimodal Signals for
Emotion Recognition in the Context of Human-Robot Interaction
This paper presents a proposal for the identification of multimodal signals for recognizing
human emotions in the context of human-robot interaction, specifically, the following
happiness, anger, surprise and neutrality. We propose to implement a multiclass classifier
based on two unimodal classifiers: one to process the input data from a video signal and
one that uses audio. On one hand, for detecting the human emotions using video data we have
multiclass image classifier based on a convolutional neural network that achieved 86.4% of
generalization accuracy for individual frames and 100% when used to detect emotions in a
stream. On the other hand, for the emotion detection using audio data we have proposed a
classifier based on several one-class classifiers, one for each emotion, achieving a
accuracy of 69.7% . The complete system shows a
generalization error of 0% and is tested with several real users in an sales-robot
Fast Path Planning Algorithm for the RoboCup
Small Size League
Plenty of work based on the Rapidly-exploring Random Trees (RRT) algorithm for path planning
time has been developed recently. This is the most used algorithm by the top research teams
in the Small
Size League of RoboCup. Nevertheless, we have concluded that other simpler alternatives show
results under these highly dynamic environments. In this work, we propose a new path
that meets all the robotic soccer challenges requirements, which has already been
implemented in the STOx’s
team for the RoboCup competition in 2013. We have evaluated the algorithm’s performance
using metrics such
as the smoothness of the paths, the traveled distance and the processing time and compared
it with the RRT algorithm’s.
The results showed improved performance over RRT when combined measures are used.
Identification of Multimodal Human-Robot
Interaction Using Combined Kernels
In this paper we propose a methodology to build multiclass classifiers for the human-robot
problem. Our solution uses kernel-based classifiers and assumes that each data type is
by a different kernel. The kernels are then combined into one single kernel that uses all
involved in the HRI process. The results on real data shows that our proposal is capable of
generalization errors due to the use of specific kernels for each data type.
Also, we show that our proposal is more robust when presented to noise in either or both
Methodology for Learning Multimodal
Instructions in the Context of Human-Robot
Interaction Using Machine Learning
This work shows the design, implementation and evaluation of a human-robot interaction
where a robot is capable of learning multimodal instructions through gestures and voice
issued by a human user.
The learning procedure can be performed in two ways: an instruction learning phase, where
the human aims
at teaching one instruction to the robot by performing several repetitions and an
instruction receiving phase
where the robot reacts to the instructions given by
the human and possibly asks for feedback from the user to strengthen the instruction’s
Design and Implementation of an Automatic Object
Recognition System using Deep Learning and an Array of One-Class SVMs.
Accepted to appear in 17th International Conference on Machine
Learning and Applications (ICMLA 2018), Orlando Florida, 2018.
Learning Soccer Drills for the Small Size
League of RoboCup
This paper shows the results of applying machine learning techniques to the problem of
plays in the Small Size League of RoboCup. We have modeled the task as a multi-class
problem by learning the plays of the STOx’s team. For this, we have created a database of
for this team’s plays and obtained key features that describe the game state during a match.
We have shown experimentally, that these features allow two learning classifiers to obtain
accuracies and that most miss-classified observations are found early on the plays.