Autonomous systems for navigation, mapping and objective completion.

This project studies how to train a robot using computer simulation, and then deploy the robot in the real world. The specific application is to train robots to use their arms and hands, to perform useful tasks by physically interacting with their environment, such as in manufacturing.
This project aims to close the gap between simulation and reality, such that a robot can be trained only in simulation, without any real-world data at all. Current approaches to this typically involve randomising parameters in the simulation, such as colours and lighting conditions, and physical properties such as mass and friction. The robot is then trained across a range of these random environments, so that when deployed in the real world, the robot is able to deal with a range of real-world conditions. However, whilst the learned robot controller is robust across different environments, it lacks precision in any one particular environment. This project will now focus on enabling the controller to adapt online to the particular environment the robot is in, such that high-precision control can be achieved in the real world.
The proposed method involves training two neural networks. The first network is trained with randomised environment parameters, as described above. The second network is then trained to adapt the output of this first network according to the observations made by the robot in its specific environment.

The main EPSRC research area for this project is "Robotics", and it also falls under "Artificial intelligence technologies" and "Manufacturing technologies".