alina tong

We're developing a novel policy training method that searches through multiple learned RL policies using MCTS to generate expert trajectories. Expert demonstrations are distilled into a point-cloud task-planner and low-level policies.

We present hybrid VLM + dependency graph-based task-planner, and demonstrate its effectiveness in open-set object retrieval in clutter. Evaluated on real and simulated environments.

• Built imitation learning infrastructure in Python and ROS Noetic to train and test robot policies on the UR5e and Motoman SDA10F robots
• Used MuJoCo to test and train robots in simulation
• Built automated object placement and scene generation tooling for domain randomization
• Built keyboard teleoperation with inverse kinematics and set up GELLO teleoperation framework
• Built a data-collection system that saves human demonstrations as LeRobot datasets
• Collected demonstrations to fine-tune the pi-0 VLA model for the UR5e

• Side-project to learn reinforcement/imitation learning by applying techniques to the fighting game Guilty Gear Strive
• Reverse engineered the compiled game code and built a C++ frontend to interact with the video game logic and a Python backend to run the AI model and control the training
• Trained a DQN model to play the game, using PyTorch and Numpy
• Scraped replay data of matches and trained a behavioral-cloned model to play the game

• Final project for CS439 to identify the efficacy of transfer learning in sentiment analysis
• Conducted literature review of sentiment analysis, learning about lexicons, Naive Bayes, parts-of-speech tagging, aspect-based sentiment analysis, sentence + document sentence-level sentiment analysis
• Scraped data from review platforms (Steam, IMDB, Metacritic) and fine-tuned the BERT models to predict review ratings from text
• Compared performance of the generalist model trained on all data against specialized models trained solely on each platform in a 7-page Latex report

• Built a simulation for an agent pathfinding through a 2D maze using variations of the A* algorithm in Rust
• Evaluated and summarized performance of algorithms in a 6-page Latex report
• Designed a two-tailed t-test for evaluating the difference between Repeated Forward A* and Repeated Backward A*

• Built bottom-up compiler IR instruction scheduler in React, NextJS, with a D3.js visualization of the instruction dependency graph, deployed on Vercel

• Built a feed-forward neural network from scratch using Rust
• Trained model to classify digits from the MNIST dataset and ascii faces, and achieved accuracy of 90.57%
• Evaluated and summarized performance of neural network in a 8-page Latex report

• Worked with a PhD student on an Imitation Learning project, building software infrastructure
• Worked with a PhD student on a VLM task-planning paper
• Presented a talk on our imitation learning infrastructure to group of 20+ students and faculty

• Worked in a team of 8 on a full-stack single-paged dashboard web-app for monitoring and analyzing fiber-optic networks, using D3, VueJS, Fastify, Prisma, and Postgres
• Built and deployed migration CLI that can create, track, and apply migrations in production databases
• Optimized line graph SQL query to run x50 faster
• Dockerized backend databases for easier local testing
• Scripted a build-system for backend API + developer tools, used in a GitHub actions CI/CD workflow
• Built a simulation CLI tool that generates 1k+ entries for spiked alarm trend DB data to demo dashboards to customers

• Led development in a team of two to build six interactive AR sites for the NJ city of Lambertville's "Flowing Together" walking tour, using AFrame and ThreeJS
• Hosted AR sites on a website built using Vuejs and Nuxt, serving 900+ monthly users