Deep Learning for Robust robot Control

Themes: Robotics, Technology & Intelligent Systems

A TRL is a measure to indicate the matureness of a developing technology. When an innovative idea is discovered it is often not directly suitable for application. Usually such novel idea is subjected to further experimentation, testing and prototyping before it can be implemented. The image below shows how to read TRL’s to categorise the innovative ideas.

Click to enlarge

Summary of the project

Robots can perfectly follow a set of commands to achieve a certain task. Up till now most of these commands are being programmed by hand. To create a robust control system for robots there is a need for a more effective learning system that can deal with unanticipated events and uncertainties in the robot’s environment. In this project the researchers aim create such control system by combining or integrate reinforcement learning (RL) strategies with deep learning strategies (DL). Reinforcement strategies determine what should be learned and the deep learning strategies model the neural network which can be used for learning. This  works well in a game environments. However, that doesn’t directly transfer to learning real-world control tasks. The research so far has shown that the decision on what to remember and what to learn from those memories  is also a fundamental aspect for more robust control. Do we want robots with a perfect memories, a memory that only remembers the last hour or should it also remember its mistakes to be able to learn from them?

What's next?

The next step is to combine this method with other learning methods for robots and with help from humans to see if the learning process can be further accelerated.

Tim de Bruin


Dr. Jens Kober