09. Learning based model estimation of robots

Our dream is a robot that learns from zero prior knowledge its capacities by exploring its actuators and sensors. Like a helpless-born human baby that must explore its own muscles and senses for learning how to use them. In robotics, a model of the dynamics of the robot and, possibly, of its environment is needed for the controller that determines the appropriate actions to be taken for performing a task. Also, for optimal interpretation of the sensor signals, a model is needed. For instance, in the context of a Kalman filter which is used to determine the robot’s state. 

We propose a data-driven self-learning approach. The robot learns by exploring and interacting with its environment. This approach was already successfully used for calibration of the robot’s sensors. At the ETRO department, we build an instrumented robot and provide a test environment. 


Sensors include the speed of the wheels, compass, IMUs (Inertial measurement Units), and a camera (for supervised position learning). 

In an industrial context, robots are used to perform increasingly complex tasks. A critical element is the controller, which determines the appropriate actions to be taken. This controller needs a model of the dynamics of the robot and, possibly, of its environment. 

In this master thesis, we propose to estimate this model from measurements, by using a data-driven self-learning approach. A balance will be sought between letting the robot learn its dynamics by itself, by using an exploration approach, or by defining informative experiments yourself. In the process, the robot will also need to learn the characteristics of its sensors and actuators. In this way, a combination of prior knowledge information and measurements from sensor will be used to execute specific tasks. 

Tentative work plan 

  • Literature study: learn about  
    • Kalman filters, self-supervised learning 
    • the control strategies used in robots and the associated required models 
    • the available techniques to extract models from prior knowledge and measured data 
  • Familiarise with the Robot Operating System (ROS) and the protocols to communicate with the robot 
  • Design experiments to extract informative data about the dynamics of the robot and the characteristics of the sensors 
  • Extract a model of the relevant elements of the robot. The model structure can be determined by using a machine learning approach or from physical knowledge (or a combination of the two) 
  • Use the model in a control strategy to let the robot perform tasks
Jan Lemeire (ETRO)
Back to top