• Completed product code pipeline transfer from Python to C++ by utilizing CMake, GDB, Valgrind• Speed up code by x2 with multithreading and achieved x3 feature matching performance

Project: Uncertainty Quantification in Language-conditioned Manipulators [1] Supervisor: Dr. Nikolai Matni• Designed model calibration approach for multiple pre-trained, language-conditioned robotic models, using PyTorch and Docker, reduced expected calibration error by 50% on average• Proposed and implemented a scalable action-selection strategy for multiple ViT-based robot manipulators, improved average success rate by 3% without any additional training on 400GBs of data for diverse tasksProject: Navigation to Multiple Semantic Targets in Novel Environments [2] Supervisor: Dr. Kostas Daniilidis• Deployed panorama sensing for the agent in Python and improved segmentation with groundedSAM• Collaborated with the team to extend existing navigation pipeline PEANUT for unordered multi-object goal navigation, provided debugging and demonstrating tools support to the team, and used Git for version control[1] B. Wu, B. Lee, B. Bucher, K. Daniilidis, N. Matni “Uncertainty Aware Deployment of Pre-trainedLanguage-Conditioned Imitation Learning Policies” (Accepted at IROS, 2024)[2] B. Bucher*, K. Ashton*, B. Wu, K. Schmeckpeper, S. Goel, N. Matni, G. Georgakis, K. Daniilidis “Unordered Navigation to Multiple Semantic Targets in Novel Environments” (CVPR Workshop, 2023)

Project: Vision Language Navigation on robot platform• Implemented machine learning pipeline in Python with PyTorch for real world robotic navigation research• Generated occupancy and semantic segmentation maps using depth projection and DeepLabv3 (PyTorch)• Coded customizable navigation action-taking commands using ROS on Roomba and turtlebot3

• Completed Wi-Fi scan with designated scan interval and wait time with RTOS in ARM Cortex-M4• Leveraged an existing C/C++ microprocessor SDK to reduce power consumption by 10%• Established SPI communication between devices and stored AP info using two space efficient data structures

• Provided after-sale services for 3D printing machines in collaboration with product engineers• Investigated quality control of printed models and product polishing including waxing and cleaning