Conducting PhD research under Prof. Sergey Levine

- Investigated use of large pre-trained language models to improve robot navigation and 3D scene understanding- Gave spotlight talk and submitted first-author paper on use of language models for zero-shot room classification at Robotics: Science and Systems 2022 - Scaling Robot Learning workshop (https://arxiv.org/abs/2206.04585)- Developed novel methods for fine-tuning language models for generalizable robot 3D scene understanding, allowing robots to recognize key room/building types- Researched use of language models to improve navigation for robot object-finding tasks in Habitat AI simulator environment- Built method for extracting common-sense spatial-semantic relations from language models, encoding said information in a spatial ontology for large-scale task-driven scene understanding

- Staffed office hours, answering student technical questions, developing homework assignments, and managing class logistics for ~380 students in MIT’s graduate-level natural language processing class- Arranged and delivered technical introductory tutorial on designing and training neural networks in PyTorch

- Reprised role for Robotics: Science and Systems as graduate teaching assistant

- Investigated methods for leveraging differentiable robot physics simulators for sample-efficient and robust learned bio-inspired legged locomotion- Designed and ran large-scale parallelized differentiation-based and reinforcement learning remote experiments on OpenMind SLURM cluster, visualizing performance metrics in Seaborn- Implemented and tested cutting-edge short-horizon actor-critic simulator differentiation-based learning algorithm

- Investigated proprioceptive sensing for mapping and navigation on Exobiology Extant Life Surveyor snake robot for exploring Enceladus’s polar water vent crevasses- Created ROS software for providing noisy IMU readings in simulation, integrated system with optimization-based kinematic inertial odometry codebase using GTSAM C++ package to leverage multiple IMUs and forward kinematics for odometry- Conducted extensive literature review to inform control policy and sensor requirements for contact-based sensing, presented results to perception and locomotion teams

- Engaged in scrum workflow, participating in daily standup briefings and biweekly development sprint meetings - Handled tickets via Jira, fixed bugs and added features to improve visualization, usability, and stability in ROS C++ codebase

- Extended MobileNet neural network architecture in PyTorch for depth and semantic learning using the virtual photorealisticBlackbird dataset, for eventual use with drone-in-the-loop experiments with the FlightGoggles simulator- Leveraged Ray Tune library for network hyperparameter search and multithreading Python library for concurrenttraining/validation data processing

- Won second place in Robotics: Science and Systems, MIT’s capstone robotics class’s annual autonomous race in my first-year, and accepted teaching assistantship- Supervised 20 teams of four or five students to complete technical software labs, briefings, and reports on robotics algorithms- Developed LIDAR-based navigation, visual servoing, path planning, and control software lab assignments on physical robot racecar and in TESSE simulator- Staffed office hours, assisted students with debugging software implementations in ROS and Python- Gave class lectures on ROS and LIDAR perception/control for wall-following- Provided feedback on technical team presentations

- Accepted teaching assistantship for MIT’s Intro to Deep Learning class, sponsored by NVIDIA and Google and taken by over 540 students- Staffed office hours on TensorFlow lab assignments on multilayer perceptrons, recurrent neural networks for music generation, computer vision with convolutional neural networks and variational autoencoders for de-biasing, and reinforcement learning for control

- Constructed a robot in a team of 5 students to compete in MIT’s annual MASLAB competition, winning second place- Built circuitry integrating NUC & Teensy with motor controllers and LIDAR, programmed localization algorithms with ROS

- Constructed software integrating NVIDIA’s IsaacSim robot simulator with RTR’s perception and control pipeline for virtual perception algorithm testing as sole end-to-end developer- Developed and documented Python software to start simulator, load robots and cameras from URDFs and calibration files, simulate real-time ROS sensor data and robot states, and enable virtual obstacle creation- Demonstrated and recorded autonomous obstacle detection and avoidance by virtual UR3 robot arms, as well as control of Dematic’s multi-robot work cell used in RTR’s palletizing service - Presented project to rest of perception team, discussed software functionality and structure, covered technical goals and challenges, and showed final demonstration videos

- Created computational model of electron transfer in novel energy storage systems using Marcus equations- Awarded Regeneron Science Talent Search semifinalist status, NYC Science and Engineering Fair finalist status, and U.S. Air Force Science Fair Program prize