Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Designing Robot Control Using Machine Learning
The automated design of a control strategy for virtual robots
Soft robots have the advantage over conventional robots that they are intrinsically compliant and have a large number of degrees of freedom. The vast design space and the complex dynamic interaction between activation and the elastic body of the robot make designing them by hand challenging, often requiring a large number of iterations. It is thus advantageous to design soft robots using a computational design approach that integrates simulation feedback. Designs are generated, evaluated in simulation and optimized. In order to evaluate designs in simulation, a control strategy is needed. For every design, a separate suitable control needs to be found. In this project, virtual soft robots are optimized for locomotion. The search space of control strategies can be reduced by limiting the search to viable solutions. Candidate methods to solve the problem are artificial neural networks, recurrent neural networks and reinforcement learning.
Soft robots have the advantage over conventional robots that they are intrinsically compliant and have a large number of degrees of freedom. The vast design space and the complex dynamic interaction between activation and the elastic body of the robot make designing them by hand challenging, often requiring a large number of iterations. It is thus advantageous to design soft robots using a computational design approach that integrates simulation feedback. Designs are generated, evaluated in simulation and optimized. In order to evaluate designs in simulation, a control strategy is needed. For every design, a separate suitable control needs to be found. In this project, virtual soft robots are optimized for locomotion. The search space of control strategies can be reduced by limiting the search to viable solutions. Candidate methods to solve the problem are artificial neural networks, recurrent neural networks and reinforcement learning.
The goal of this thesis is the automated design of a control strategy for virtual robots. Deliverables are C++ code that finds a suitable control strategy for a variety of robotic designs in a feasible time and a strategy to limit the search space to consist of mostly viable solutions, but still containing a large variety of solutions.
The goal of this thesis is the automated design of a control strategy for virtual robots. Deliverables are C++ code that finds a suitable control strategy for a variety of robotic designs in a feasible time and a strategy to limit the search space to consist of mostly viable solutions, but still containing a large variety of solutions.