A low cost motion capture system for improving the reality of humanoid robot simulator by machine learning

Abstract

Simulation is very important tool in the development of robot system. However it is really difficult to build a simulator which behaviors close to real robot, especially in the domain of humanoid robot motion, e.g. biped walking. This is because there are a lot of parameters in simulator needs to be adjusted according to real system, and some mechanisms are difficult to be modeled explicitly. Machine learning algorithms can be used to optimize parameters or learning implicit models, but it needs a lot of training data. Traditional motion capture system are expensive and limited for this. In this paper, we use a low cost RGB-D sensor (e.g. kinect) to collect motion data of humanoid robot NAO, and use them as training data of machine learning for improving the reality of RoboCup official simulator -- SimSpark.

@Conference{lc3d,
  Title                    = {A low cost motion capture system for improving the reality of humanoid robot simulator by machine learning},
  Author                   = {Yuan Xu and Hans-Dieter Burkhard},
  Booktitle                = {Workshop: Low-Cost 3D – Sensors, Algorithms, Applications},
  Year                     = {2011},

  Address                  = {Berlin, Germany},
  Month                    = {December},

  Abstract                 = {Simulation is very important tool in the development of robot  system. However it is really difficult to build a simulator which  behaviors close to real robot, especially in the domain of humanoid robot motion, e.g. biped walking. This is because there are a lot of parameters in simulator needs to be adjusted according to real system, and some mechanisms are difficult to be modeled explicitly. Machine learning algorithms can be used to optimize parameters or learning implicit models, but it needs a lot of training data. Traditional motion capture system are expensive and limited for this. In this paper, we use a low cost RGB-D sensor (e.g. kinect) to collect motion data of humanoid robot NAO, and use them as training data of machine learning for improving the reality of RoboCup official simulator -- SimSpark.},
  Owner                    = {xu},
  Timestamp                = {2012.02.01}
}
Autoren:
Yuan Xu, Hans-Dieter Burkhard
Kategorie:
Tagungsbeitrag
Jahr:
2011
Ort:
Workshop: Low-Cost 3D – Sensors, Algorithms, Applications