The development of a universal robot control system using limited feedback, implemented using the Fido library.
The Fido project originally started as a universal robot control system, created as part of a Science Research class to address the following goals:
These goals were achieved through the training of artificial neural networks with a wire-fitted moving least squares interpolator following the Q-learning reinforcement algorithm and an action selection policy that utilizes a Boltzmann distribution of probability, all implemented using the Fido library.
The control system was initially tested using the Fido simulator, allowing rapid prototyping and iteration. Next, a small robot using the Intel Edison compute module was constructed for physical testing. The robot was successfully trained to do a variety of tasks with limited feedback, such as staying put, moving only in the dark, and driving to a point. Reward was given to the robot over wifi through an app developed in the Ionic framework.
Another hardware implementation, this time implemented on a Raspberry Pi Zero, is currently in development.
The paper, research poster, and other multimedia can be found in the linked Github repository.