Projects
Simulations
In my observation, the rise of Gaussian Splatting (GS) is setting the stage for a revolution in 3D reconstruction and virtual testbeds for digital twins, leaving conventional methods like photogrammetry and Neural Radiance Fields (NeRFs) in the dust. I recently had the chance to test it out —by taking a short, 10-second aerial video of the stunning landscape around Blüemlisalphütte, Switzerland. From that brief clip, I created a rough Gaussian splat, reconstructed the terrain in Gazebo, and then implemented Region of Interest (RoI) detection using a simulated drone in ROS Gazebo.
More Details:
I performed a 3D map simulation of V-shaped Martian cave using a 3D LiDAR mount on my designed robot, Big bot. Octomap was used as the mapping approach. It was featured on ROS News for the Week, January 23rd | accessible at - ROS News for the Week of January 23rd, 2022 - General - ROS Discourse
More Details:
A test of autonomous SLAM and navigation to be implemented in my designed nurse robot. According to preliminary simulations, the bot is capable of transporting specified medical products, such as pharmaceuticals, and performing rudimentary medical procedures at a location using built-in algorithms.
More Details:
I developed an affordable conceptual search and rescue robot that can detect victims in disaster rubble using affordable 2D LiDAR instead of expensive cameras and also create a 3D map of the disaster site for further investigation. This was developed as part of a project showcasing competition arranged by IEEE in 2021.
Further reading:
The first suspension test of the design of our mars rover prototype, Ogrodoot for IRDC 2021. We made some major modifications to the wheels after the test. This was by far the most stressful project we worked on this year having simultaneous lab finals and end-semester tests.
Extremely proud of my team, Ogrodoot for pulling off such a big stunt in half a week against the tide.
Further reading:
-'Cause I'm mostly broke.
Physical Implementations
I used hector_mapping from the hector_slam package as it can work fairly well without odometry data. There were some initial hiccups (I guess you might call them the perils you encounter when you step into reality from simulation!) but it all worked out very nicely in the end, as evidenced by the 2D-map.