Terrain Traversing Robots
- Technologies -
TensorFlow | Machine Learning | Unity
- Skills -
CAD | Optimisation | Research | Code
Why?
The evolution of computers has led to the development of powerful machines that can automate various aspects of society.
However, the effectiveness of human-controlled robots is limited by factors such as reaction time, environmental information,
and control limitations.
Machine learning can be used to improve the navigation abilities of robots. The aim of this project
was to design and train simulated robots of varying complexity to observe how they navigated a given environment and identify
real-world applications for these designs.
The objectives included building code that dictated agent behavior, creating
environments with some level of randomness to ensure the agent was actually learning, and logging and comparing the
learning patterns and training time of the agent.
Visualisation
Summary
The tested designs could be used in various real-world applications, such as exploration and search and rescue.
The sphere design could be mass produced and used as a basis for nanobots that can connect with magnets and create
temporary structures.
The conjoined capsule design, with its narrow frame and joints, could be used to access tight spaces
and climb.
The 4 and 8 limbed models could be used for carrying luggage or weight and may have practical applications in fields
such as the military.
It was found that a centralised brain was more effective and that improving the script of the agent brain
was more beneficial than longer training sessions.
The project's unity source code will be open sourced on GitHub once it has been refined.