Harry Freeman

I am a PhD student in the Robotics Institute at Carnegie Mellon University advised by Professor George Kantor Carnegie Mellon University in the Kantor Lab.

I received my B.S. in ECE from Cornell University and my Masters in Robotics at CMU, where I worked on computer vision, 3D reconstruction, and next-best-view planning to phenotype small crops in agriculture. Prior to CMU, I was a senior embedded software engineer for Amazon Web Services in their AI Devices division working on AWS Panorama.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Github

My research interests lie at the intersection of 3D reconstruction, robotic manipulation, and learning from human demonstration. I wish to enable robots to perform complex tasks in diverse and unstructured environments such as agriculture. I am currently working on cross-embodiment learning from human demonstration utilizing gaussian splatting.

WARPED: A framework that synthesizes realistic wrist-view observations and actions from egocentric human demonstration videos to facilitate the training of visuomotor policies using only a single monocular RGB camera.

Created a method for temporal apple fruitlet association utilizing stereo images and transformers. Able to achieve F1 matching accuracy of 92.4% on new dataset collected over 3 years of 3 different varietals. We demonstrate that our transformer architecture is generalizable to other datasets and modalities.

Developed a novel next-best-view planning approach to enable a 7 DoF robotic arm to autonomously capture images of apple fruitlets. Utilized a coarse and fine dual-map representation along with an attention-guided information gain formulation to determine the next best camera pose. Presented a robust estimation and graph clustering approach to associate fruit detections across images in the presence of wind and sensor error.

We develop a method for creating high-quality 3D models of sorghum panicles to non-destructively estimate seed counts. This is acheived using seeds as semantic 3D landmarks for global registration and a novel density-based clustering approach. Additionally, we present an unsupervised metric to assess point cloud reconstruction quality in the absence of ground truth.

We demonstrate how the use of semantics and environmental priors can help in constructing accurate 3D maps for downstream agricultural tasks with the target application of phenotyping Sorghum.

We build a real-time aerial system for multi-camera control that can reconstruct human motions in natural environments without the use of special-purpose markers. This is acheived with a multi-robot coordination scheme that maintains the optimal flight formation for target reconstruction quality amongst obstacles.