392262 ISY Project: Learning to Walk - a Biologically Inspired Learning Approach for a Six-Legged Robot (Pj) (SoSe 2020)

Animals excel at adaptive behavior and can navigate quite difficult terrain, for example, climbing through twigs as found in stick insects. While current Deep Reinforcement Learning (DRL) approaches have recently shown to provide solutions for robots in well defined scenarios and tasks, these don’t scale well towards unpredictable environments and when requiring fast adaptations. Here, we bring together insights from biological research on motor control in insects and DRL for the task of six-legged walking on a small robot. A decentralized control system as found in stick insects will be used as an architecture for DRL of locomotion.

Currently, there is a detailed simulation of a six-legged robot controlled through ROS. The tasks for the group will be to connect this to the existing robot. First, providing the connection to ROS. Second, implement a baseline implementation of a simple biological control architecture. Third, adapt the learning architecture in a way that can be employed on the real robot.

In case this would not find enough interest for a team project, this project proposal would be also offered (in reduced/modified form)
[x] as individual project 
[x] as project for 2-3 students

Requirements for participation, required level

Introduction to Neural Networks and python programming.
It would be good to have some students with experience in ROS and some with basic knowledge on DRL.