Xserver G5 Cluster

Developer Profile: INNlink.

This Danish Technical University based AI start-up company uses Unity to train their next-generation neural-networks. Autonomous bots are simulated in a virtual environment, using sensors to probe their surrounding. The simulations are set to evolve on a large cluster of Apple Xserve G5s.

What is INNlink?

INNlink consists of Christian Bentzen, Allan Hammershøj and Emil Heinze who are masters of Science in engineering from the Technical University of Denmark. Each of us have knowledge in different areas of science, and what connects us is our common interest in artificial intelligence. We have each chosen a different approach to this topic, and we seek to combine our differences to create products that are much more intelligent than the standard of today.

With expertise in classic disciplines in artificial intelligence especially in fuzzy logic, INNlink specializes in Spiking Neural Networks. We use results obtained from physical measurements of living brain tissue to model the processes that goes on in living brains. Through this work we focus on building new navigation units for robots inspired by the human brain and auxiliary equipment for disabled persons such as the hard of hearing, blind people etc. The technology, however, has endless possibilities. Through project work in collaboration with companies and Master of Science students we try to cover the many aspects of this exciting technology.

What is the project about?

INNlink has created a generic, socket based framework we call INNsim which interfaces with Unity to facilitate development of the intelligent systems. Unity is used to provide different virtual testing scenarios and environments for the INNsim framework. Through INNsim cluster computing, genetic algorithms, neural network models, virtual environments, and simulated sensor equipment are combined and an experimental platform is obtained. On this the experimenter can choose a task that a control unit should try to solve, using e.g. an evolutionary strategy Such an experiment could for example be a 3D reconstruction of a production plant, and the task could be that of an assembly or cleaning robot.

How did you work on the project?

Creating an artificial world and the necessary scripts to make INNsim and Unity communicate took two programmers just two weekends to finish.

What did that achieve?

A small virtual hovercraft with simulated spiking neurons came to life in the artificial world. The aim of the little hovercraft was to move around in the virtual world and avoid hitting things. The neural network was not trained yet.

What was the best part of the Unity experience?

The easiness of interfacing with our own system through scripting and the user interface when developing a virtual environment make Unity quite unique.

What obstacles did you run into?

A number of bugs. But most of these are now fixed thanks to the great feedback system. We also experienced problems with implementing sensors such as ray casting in Unity. The software provides very sparse feedback from the 3D engine making this task hard to implement for the programmers.