Offroad Navigation

Overview

#

Self-driving cars have come a long way from basic lane keeping assist to full autonomous capability, due to the success of Tesla and Waymo. They are pretty good at driving in urban settings, but fail miserably in offroad conditions. This is due to the absence of a structured environment, which self-driving cars take advantage of in cities.

My work as an intern at Moon Lab during the Summer of 2025, focuses on testing the offroad navigation capabilities of existing methods. I implemented a visuo-motor policy inspired from the Learning by Cheating paper, and deployed it in an ATV. The project is discussed in detail in the following sections.

TB;DW

#

(too boring, didn’t write)

my contributions:

• made docker containers to run camera and lidar on nvidia jetson
• wrote data collection scripts for carla
• implemented a resnet architecture for predicting waypoints from images
• wrote ros nodes for inference and pid control (meant to be used with carla-ros)
• ported carla to mac

resulted in:

• a blog about porting carla to mac: https://nveshaan.github.io/blog/macarla/
• a github discussion (top #1): https://github.com/carla-simulator/carla/discussions/9037
• two amazing repos:

and

what i learnt:

• ros, docker, bash, linux
• carla, pid
• pytorch, diffusion