Teaching robots how to be useful

- One of the youngest technical leads. Led camera blindness detection and environmental sensing - negotiated features and requirements with stakeholders, planned and assigned work deliverables to the team, worked across teams to meet functional use-cases, held accountable for the final delivery of production blindness detection and environmental sensing solution.- Delivered a production quality, state of the art, safety certified camera blindness detection system with limited time and resources by effectively prioritizing critical tasks. In production 2025.- Built the full stack production deep learning pipeline from data curation to model training, quantization, evaluation, and efficient in-car deployment for blindness detection.- Trained and released several production quality large-scale deep learning models for camera blindness detection and environmental sensing with a limited compute budget.- Developed a novel method for training blindness detection DNN which allowed the system to exceed prior capabilities - detecting blockages faster and more completely.- Primary author for the training pipeline of an initial 3D lane detection DNN. Had an initial prototype working after just 3 days and the full pipeline merged in 2 weeks. The training pipeline was widely re-used by many upcoming 3D detection networks. Was one of two engineers running major training experiments to improve the model performance and generalization.- Designed and implemented initial ML infrastructure for training the next-generation driving network.- Wrote safety-compliant and efficient C++ and CUDA code for the deployment and postprocessing of neural network outputs.- Designed, implemented, and trained a new dashmark detection DNN head - patent pending- Highlighted as one of the top 5 contributors in the full perception organization (100+ contributors) based on quantity of code committed, and consistently ranked as a Top Contributor (highest ranking) every year of employment.

- Built automation infrastructure to streamline DNN model development, KPI computation, and deployment- Built a system for maintaining traceability from training jobs to the code and commands used to generate them in order to improve repeatability of experiments- Designed a method to determine whether an upcoming region of road would be consistently illuminated - patent pending - Delivered a tool to mine data based on several metrics. The tool could be used for finding poorly labeled training examples or to mine challenging data from an unlabeled dataset.- Developed a new paradigm for image blindness classification which focused on actionable outputs for system deactivation.- Trained and released deep neural networks for camera blindness detection.

- Developed several methods for monitoring the outputs of neural networks for direct use in pedestrian detection monitoring. The resulting paper was accepted to IEEE Intelligent Transportation Systems Conference 2020. - Collaborated with other researchers to develop a system for building a Dynamic Scene Graph (DSG) of a building automatically, the first system of its kind. The resulting paper was accepted into Robotics Science an Systems (RSS) 2020, and it will be presented as a workshop paper at Computer Vision and Pattern Recognition (CVPR) 2020.- Developed a first attempt at verifying a full mesh of the environment by predicting per-face error using a graph neural network which operates on the dual of the environment mesh.

- Developed a sensitivity analysis pipeline to assess neural network weaknesses- Helped improve path perception neural networks through the addition of augmentations and some minor architecture changes.

• Collaborated with two postdoctorates to develop a pedestrian trajectory prediction model that integrates information from the environment• Integrated a model architecture from a top paper in the field• Refactored codebase for clarity and efficiency

• Fixed, refactored, and wrote test cases for probability distribution functions for sensor detections• Researched and developed a novel algorithm for road barrier detection and tracking using only radar data• Collaborated with other interns to develop an end-to-end indoor beacon localization system and demonstrated a parking maneuver on the real vehicle without GPS

Fall 2017 and Winter 2017-2018• Developed Deep Reinforcement Learning platform for learning autonomous behaviors. Platform can run arbitrary simulations to get data and train Deep Neural Network models. Focus on usability. All aspects of the neural net architecture as well as state and action vector definitions can be modified by changing one file. • Developed new class to embed python interpreter in C++ script to call Python Keras libraries from the learned behavior written in C++. • Resulting platform allows for a new process of behavior creation using Deep Reinforcement Learning to make more complicated behaviors that can learn intelligent interactions between robots. • Code publicly available at: https://github.com/argupta98/moos-ivp-pLearn• Wrote a conference paper on the integration of Reinforcement Learning with MOOS-IvP, and will be presenting the paper at OCEANS conference in October 2018.Summer 2017• Wrote two new advanced marine robot autonomy behaviors, vastly upgrading robot capabilities for teamwork and defense. New behaviors are now used for research on human-robot trust. Worked in the context of a large codebase using SVN workflow. • Trained a linear model to classify whether a robot it attacking or defending. Feature vector was a trail of x,y points and team color. Resulting classifier was 99.8% accurate. • Code publicly available at: https://github.com/argupta98/moos-ivp-internship

• Collaborated with other researchers using Git workflow to write an interpreter which converted App Inventor Block Code to code runnable on Arduino101. Built key functions to run the interpreter and call device methods. Result was the first step in allowing students to code Arduinos directly from the App Inventor environment.

• Designed and built a remote control car that could autonomously navigate and drive using Radio-Frequency(RF) beacons for indoor localization. Implemented a waypoint following algorithm on Arduino to control servos; Delivered a car that could autonomously travel to waypoints using an RF tag for localization, however, high latency in the RF technology limited car mobility