In my experience the Tegra K1 is rather limited as a deep learning platform. A Tegra X1 would easily fit in this platform, and is about twice as powerful (at about twice the cost, admittedly). In fact, for a slow-moving ground robot such as this I would consider a laptop with one of the mobile Nvidia GPUs, such as the GT960M. This would be more expensive of course (just the laptop would probably be more than $1k) but it'd be far easier to work with.