Blog

Data Problems The limiting factor of this project has been the data collection process. We created the data collection pipeline but built in some crippling factors. These mainly have to do with the format of our data. To explain these problems let’s look at the data layout: [Read More]

Looking at the gifs on our project homepage, it’s pretty clear that our robot has attempted to emulate human reflexes. In this blog post, we take a closer look, and try to understand why this behavior came to be. [Read More]

Overview We trained a lot of models - probably somewhere ~50 different ones. We compared a variety of different ofr one common dataset. The data for the corresponding models can be seen in the following figure: [Read More]

Neural Net Research One of the potential limitations of our first model is that it only looks at the current positions of the balls in the world. This is a tricky problem, most simple Neural Net models are made to perform on a single timestep’s information. [Read More]

To start off our machine-learning-robo-dodgeball adventure, we first needed a dataset to train our glorious neural net with - and what better way to do that than to collect it in our favorite ROS-compatible simulator, Gazebo! [Read More]