Ryan works with PhD student Radha Lahoti on a differentiable deformable simulator and reduced-order modeling. His research helps make physics simulations more efficient and trainable for robotics. (Ryan 2/3)

This week we’re featuring Ryan Chaiyakul, a graduating UCLA junior in Computer Engineering from San Diego. He’s passionate about robotics, reinforcement learning, and embodied models - especially where control meets perception. (Ryan 1/3)

Henry Braid is working with Mason Zhao and Shivam Kumar Panda on how LLMs can improve neural network workflows for physics-based simulations. Using a ribbon model of elastic deformation, they explore LLM-guided optimization for better performance. (Henry 2/2)

This concludes our 3D Reconstruction post series. We hope to share the publication and more details soon. (Reconstruction 13/13)

This week we’re featuring Henry Braid, a UCLA senior majoring in Math of Computation with a Data Science minor. He’s passionate about software engineering, ML, basketball, and making music on the piano and guitar. (Henry 1/2)

Experiments show NIRSplat outperforms InstantSplat and traditional 3D Gaussian Splatting across 3, 6, and 12-view setups. It achieves higher SSIM, PSNR, and LPIPS scores, proving its strength in complex agricultural environments. (Reconstruction 12/13)

Using a mobile robot (Scout 2.0), Jaehwan collects field data and develops real-time navigation that adapts to crop density. From hardware design to 3D reconstruction, his full-stack work pushes toward smarter, more efficient agriculture. (Jaehwan 3/3)

Our results show that botanical metadata - NDVI, NDWI, chlorophyll index, temperature, and humidity - plays a key role in accurate 3D reconstruction. NIRSplat leverages this context to boost performance and enable robust, reliable modeling. (Reconstruction 11/13)

NIRSplat is our transformer-based multimodal architecture that fuses RGB, NIR, and botanical metadata. With cross-attention and 3D positional encoding, it achieves high detail and accuracy, even under occlusion or variable lighting. (Reconstruction 10/13)

Jaehwan is working on 3D farm reconstruction using unmanned ground vehicles (UGVs). His project helps farmers monitor crop conditions remotely through real-time, interactive 3D scenes, leveraging techniques like SLAM and 3D Gaussian Splatting. (Jaehwan 2/3)

Our setup captures plants from 24 viewpoints during full 360° rotations, ensuring precise multi-view coverage. Adjustable sensor height and consistent indoor/outdoor procedures maintain alignment and reliability under varied lighting. (Reconstruction 9/13)

This week we’re featuring Jaehwan Jeong, a Visiting Grad Researcher at UCLA MAE from Seoul. He’s passionate about combining computer vision and robotics for real-world impact from autonomous nav to smart agriculture. (Jaehwan 1/3)

NTRPlant is our multimodal dataset for agricultural 3D reconstruction, combining RGB, NIR, depth, LiDAR, and metadata. It addresses lighting variation and complex plant structures, offering more diversity and coverage than existing datasets. (Reconstruction 8/13)

We use DSLR, LiDAR, depth, and NIR data plus temporal and farm metadata to boost 3D reconstruction in agriculture. Our goal is to build high-fidelity models for plant health and precision farming by refining state-of-the-art methods. (Reconstruction 7/13)

Adam’s experiments use Shrinky Dinks and Kirigami-inspired designs to create bistable structures for future robotics. The spoon shown here was made from a flat 2D layer that transformed into 3D after heat—no assembly required. (Adam 3/3).

Our goal is to build a robust AI system for 3D reconstruction in agriculture. We tackle real-world challenges like lighting variation, object motion, and limited plant datasets by evaluating methods like NeRF, Mip-NeRF, and TensorRF in farm environments (Reconstruction 6/13)

Adam is working with postdoctoral scholar Dr. Yupeng Zhang on “Machine Learning Assisted Physics-Based Modelling of Bistable Composites.” His research blends computational modeling with hands-on prototyping to understand materials that can snap between stable shapes. (Adam 2/3)

Yupeng (postdoc) and Siwoong (Samsung scientist) are building a fast, physics-based deep-learning model for composite plates—plus reverse-engineering 2D prints to create 3D morphing structures. Think smart actuators for next-gen robots. (Project 10/10)

This project is supported by the USDA NIFA grant “DSFAS: AI-based Fast Automated 3D Scene Reconstruction of Plants and Farms for Agricultural Robots,” in collaboration with North Dakota State University (Award #2024-67021-42528. (Reconstruction 5/13)

SCI Lab’s Scout 2.0 rolls through the fields with stereo + NIR cameras, 3D LiDAR, and swappable spectral sensors. Paired with AgroNav, it autonomously maps farms in high-res 3D, building time-lapse digital twins for smarter crops. (Project 9/10)