Contents
Robotics
Publicly Available Videos:
- Terrain Map Conditioned Autonomous Earthmoving Early Model Test
- Autonomous Earthmoving Evals
- Map-Assisted Terrain Smoothing with the REO-RCTL
Publicly Available Writeups:
- Terrain Map Conditioned VLAs for Autonomous Earthmoving
- A Robust Localization Solution for an Uncrewed Ground Vehicle in Unstructured Outdoor GNSS-Denied Environments
Web Articles:
- ERDC robotics team advances future of combat engineering at PC-C5
- Unmanned Ground Vehicles successfully demonstrated at PNTAX ‘23 (jammed + night!)
- S&T Award
Last Edited 04/24/26
Low Fidelity Dozer RL Environment
In Short: Unoptimized, graphics are ok , still some physics to work through. WIP.
This had been dormant for a while, but after reading this paper from ETH Zurich, we came back to the project. The problem is a bit different, since vehicle locomotion is actually necessary for earthmoving itself. Which means the entire vehicle has to be modeled to some degree, necessitating a 3d representation, all being a fair bit computationally intensive. At least relative to other projects using pufferlib.
Some more details. So I haven't actually started combing for optimizations. I just wrapped up the kinematics, and am wrappign up the physics. These's still some issues with backdragging, so I have to review cells are iterated for the soil interaction portions of the sim. Also, the dozer is just a series of floating boxes (representing joints) right now. All the global joint poositions and orientations are calculated in the sim. For actual use in RL I'll just make this a runtime flag or something, so as to only calculate relative to global when running the visualization.
But, it's semi-realistic kinda. Ideally witth some improvements and scoped domain randomization, a decent enough policy can be trained.
Last Edited 06/02/26
Microstrain ROS2 Drivers
Construction-grade IMUs in ROS2!
Last Edited 11/07/25
Visual Pose Estimation for Bulldozer Implement
Stereo camera implementation of blade pose tracking. Developed mainly for a project for my 3D vision class. As a result it's a combination of a few different 3D vision techniques. Stereo cameras for depth, removing the background, getting keypoints, and essentially doing 3D reconstruction on those points but in reverse, with the camera stationary and the objects moving.
Last Edited 11/07/25