Aaron Huang Email & Phone Number

Foster City, California, Us

Various projects working on reinforcement/imitation/value learning and multimodal foundation models

Pittsburgh, Pa, Us

- Researched deep reinforcement learning and imitation learning methods for self-driving cars in simulation- Created a failure analysis and visualization toolbox for autonomous driving performance on the CARLA Leaderboard benchmark- Developed an autonomous driving stack on a small-scale MIT RACECAR testbed that comes with LiDAR, stereo cameras, and a Jetson TX2 embedded computer

Foster City, California, Us

Used machine learning to predict interaction graphs for planning and decision-making in autonomous vehicles

San Francisco, California (Ca), Us

- Used PyTorch to design and train convolutional neural networks for real-time drivable path detection and object detection. Deployed models on the production self-driving stack as ROS nodes- Designed a CNN training procedure that jointly trained drivable path and object detection models with a semi-supervised composite loss function. Explored RNN models for better object/drivable path tracking- Trained models on an AWS cluster and worked with Django databases with over 70K image annotations. Adjusted CNN training workflows to achieve a 3x reduction in training time

Cambridge, Ma, Us

- Studied robotic navigation through structured, unknown environments using novelty detection and deep learning to identify situations where machine learning models should not be used- Implemented a state-of-the-art autoencoder approach for novelty detection in TensorFlow and trained the model to identify input images that are outside of the training distribution- Prototyped an alternate approach for novelty detection using Generative Adversarial Networks (GANs)

Cambridge, Ma, Us

- Developed a heuristic-based method in CommonLisp for fast multi-vehicle coordination with direct application to a proposed NASA lander and undersea explorer mission to Europa- Invented Risk-Aware Decentralized Model Predictive Control (RADMPC) to rapidly approximate the brute-force solution and decouple the computationally expensive problem into tractable subproblems- Achieved a 46x speedup in computation time compared to the state-of-the-art solution

Manassas, Va, Us

- Researched deep learning for aerial hazard detection, aircraft tracking, and operator awareness systems- Prototyped CNN solutions for internal trade studies and testing with deployment hardware including the Intel Movidius Neural Compute Stick (NCS) and an Nvidia GTX 1080 GPU- Used the Keras and TensorFlow frameworks with transfer learning and data augmentation techniques to train models for robust, real time performance on research and development tasks

Manassas, Va, Us

- Designed a system for optimal task allocation on the Traveling Salesman Problem as foundation for optimal task allocation for teams of cooperative UAVs- Implemented a branch and bound algorithm on a decision tree representation of the Traveling Salesman Problem using a greedy search as the initial solution algorithm- Abstracted objectives into subtask representations, allowing users to simply pass in objectives to be completed and retrieve subtask-vehicle couplings that optimally achieve objectives

Pasadena, Ca, Us

- Developed a high-fidelity, closed-loop hardware simulation for the Landing Vision System (LVS) used by NASA’s Mars 2020 rover mission for computer vision guided Entry, Descent, and Landing (EDL)- Designed models for LVS hardware peripherals that emulate the hardware down to register access, bus control, and peripheral interface levels- Integrated existing open-loop hardware models written in C into development of the closed-loop C++ simulation to emulate data flow from inertial measurement units and cameras to onboard processors

Cambridge, Ma, Us

- Assisted in testing and development of control algorithms for computer vision aided autonomous navigation and docking of the SPHERES satellites aboard the ISS- Transcribed live ISS test sessions, brainstormed improvements to existing localization algorithms, assisted glass table testing- Automated testing and analysis of a closed-loop MATLAB simulation to resolve discrepancies between simulation behavior and ISS testing behavior

Santa Cruz, Ca, Us

- Worked with a PhD candidate on his work detecting extragalactic outflow around galaxies- Created a Python script to filter and co-add light spectra data from a data pool of >2000 galaxies drawn from the DEEP2 survey- Retrieved the first generalizable instance of extragalactic outflow on a statistically significant set of data- Coauthored a research paper for the Siemens research competition. Advanced to the regional final round and was flown down to the California Institute of Technology to present and defend research against PhD candidates and professors

Bachelor Of Science - Bs, Aerospace Engineering And Autonomous Systems

Master Of Science - Ms, Robotics

High School Diploma