Suryaprakash Senthil Kumar

I am an M.S. Robotics graduate from Georgia Tech, specializing in Artificial Intelligence (AI), Perception, and Mechanics. Currently, I am a Research Affiliate in the BrainML lab under Dr. Anqi Wu , where I focus on designing learning-based motion prediction algorithms to enhance robot navigation. I received my bachelor's degree in Mechanical Engineering at SSNCE, where my thesis involved developing a quadrotor to detect and localize coastal debris.

I interned at Swaayatt Robots in Bhopal during my undergraduate studies, where I contributed to software development for autonomous vehicles. Additionally, I served as a Graduate Teaching Assistant at Georgia Tech for CS 6601/3600: Introduction to Artificial Intelligence.

I am deeply passionate about robotics, autonomous systems, and leveraging deep learning for real-time applications. Outside of work, I enjoy playing volleyball, reading, and discussing latest developments in Autonomous Driving.

If you're short on time, feel free to check out my resume for a quick overview of my work!

Email  /  Resume  /  Github  /  Linkedin

profile photo

Projects
Distributed Advantage Actor Critic (DAAC)
[Code] / [Report]

  • Developed a distributed approach facilitating model training across multiple nodes in a topology by providing mechanisms for remote communication using RPC.
  • Enhanced data collection by leveraging A3C and PAAC networks to gather experiences from a distributed cluster, achieving a 1.5X speedup over the serial counterpart.

    		•
    		Observing emergence of collaboration in a multi-agent environment
    [Code]

  • Simulated and studied the emergence of collaborative behaviour amongst players in a game of basketball in 2v2 setting.
  • Built a 2D simulator and prepared an interface compatible with Gym and PettingZoo, trained with Multi-Agent PPO.
  • Identified collaborative behavior and its strong correlation to winning by comparing the pass/win ratio of each team.

    		•
    		Human Motion Prediction: With great power comes great res-pose-ability
    [Code] / [Report]

  • Investigated sequential models for human pose prediction on computationally constrained systems.
  • Developed Temporal Convolutional Networks (TCNs) that outperformed Vanilla Transformers by 53% with less than 50% of the parameters. Achieved a 56% improvement over Transformers using a generative model with TCNs as the backbone.

    		• Obstacle avoiding and sign following mobile robot

  • Engineered a mobile robot via state machines to classify signs and avoid obstacles by fusing LiDAR, Pi-Cam & odometry.
  • Deployed motion planning algorithms, including A* and RRT* using ROS2 Nav-Stack for real-time autonomous navigation within a maze environment. Implemented pure pursuit controller for real-time lane following.

    		•
    		Painting with an Industrial Robotic Arm
    [Report]

  • Programmed a Universal Robot (UR5) to paint an outline of an object using visual cues from an image in real time.
  • Performed Forward and Reverse displacement analysis using the DH parameters and simulated robot motion using waypoint generation techniques in MATLAB and Gazebo.

    		•

    Publications
    		Multi-class segmentation of coastal debris using encoder-decoder architectures
    [Paper]

    Surya Prakash, S., Vengadesh, V., Vignesh, M., Gopal, S.K,
    Machine Learning Techniques for Smart City Applications: Trends and Solutions, Springer 2022

    Experience
    		Research Affiliate | BrainML Laboratory, Georgia Tech
    Aug 2024 - Present

    • Modeled undirected spatio-temporal graphs to analyze behavioral interactions among mice by predicting their motion.
      • Setup a distributed framework for training sequential models, with transformer-based attention to estimate edge weights. Constructed a custom loss function to preserve mice anatomy, achieving 12% improvement in final trajectory prediction.

    		Student Researcher | CORE Robotics Laboratory, Georgia Tech
    May 2023 - May 2024

    • Designed a simulation environment in MuJoCo to test an autonomous wheeled robot for mobile navigation, focusing on various control strategies using reinforcement learning.
    • Used off-policy learning & achieved around 98% success rate on mobile navigation and reacher tasks with < 1% error.
    		Computer Vision and Deep Learning Intern | Swaayatt Robots Pvt. Ltd
    Oct 2021 - Mar 2022

    • Researched CNN pruning techniques for object detection and tracking on image data from an L4 autonomous vehicle.
    • Investigated & optimized various neural network architectures, leveraging Google TPUs for large-scale training.
    • Employed Lottery Ticket Hypothesis and compressed VGG16 by 45% & YOLOv3 by 30% with 5% accuracy drop.

    This website is based on Jon Barron's personal webpage. Please free to check out the source code