Shanthosh Raaj

Shanthosh Raaj
Mohanram Mageswari

MEng Robotics · University of Maryland College Park

Researching robot learning, tactile sensing, and perception driven manipulation. Building systems that let robots feel, see, and act in the real world.

Robot Learning Tactile Sensing Computer Vision VLM Models VLA Models
01

About

I'm a Master's student in Robotics at the University of Maryland, College Park, working as a Research Assistant in the Perception & Robotics Group (PRG) under Prof. Yiannis Aloimonos.

My research focuses on building robots that can manipulate objects with human like dexterity combining vision, touch, and language to create richer, more adaptive robot behaviors.

Before graduate school, I founded a software startup (ApexMind Tech Solutions) and hold a B.E. in Computer Science Engineering from Panimalar Engineering College, Chennai, Tamil Nadu, India.

1
Paper Submitted (CoRL 2026)
15+
Robotics Projects
PRG
Lab @ UMD
02

Education

University of Maryland, College Park
Master of Engineering in Robotics
Aug 2025 – Present · College Park, MD, USA
Panimalar Engineering College
B.E. in Computer Science Engineering
Nov 2021 – Apr 2025 · Chennai, India
03

Experience

Research Assistant
Perception & Robotics Group (PRG) · University of Maryland
Feb 2026 – Present
  • Researching multi-modal representation learning for contact-rich manipulation — combining vision, tactile sensing, proprioception, and language.
  • Developing robot learning policies on a real UR5e using Diffusion Policy trained on synchronized multimodal demonstrations with GelSight Mini tactile feedback.
  • Investigating tactile-driven manipulation to improve contact awareness, grasp robustness, and execution reliability on physical hardware.
  • Building end-to-end data collection and evaluation workflows including robot calibration and policy benchmarking on task success and robustness metrics.
Founder & Software Developer
ApexMind Tech Solutions
Aug 2024 – May 2025
  • Founded and led a software startup serving small businesses, delivering custom websites and billing/invoicing systems.
  • Built full-stack solutions using Java and Python — customer records, invoice generation, payment tracking, and sales reports.
  • Led and mentored a small team, setting coding standards and running code reviews to improve delivery cadence.
Java Full Stack Intern
BrainOvision Solutions Pvt. Ltd.
Jun – Jul 2023
  • Collaborated with senior developers to build and test full-stack applications using Java, React, and SQL.
  • Gained hands-on experience in debugging, deployment, and integrating backend services with frontend interfaces.
04

Publications

CoRL 2026 · Submitted May 2026
What to Attend, What to Keep: Multi-Modal Representation Learning for Contact-Rich Manipulation

Developed a UR5e-based robot learning system combining RGB perception, GelSight Mini tactile feedback, proprioceptive state, and continuous action trajectories. Designed modality-specialized encoders, asymmetric temporal horizons, discriminative embeddings, and event-based memory for language-guided, phase-adaptive sensor prioritization. Evaluated vision-only vs. vision+tactile policy variants on physical robot hardware.

05

Projects

Map-Adaptive Informed RRT* Path Planning
Implemented RRT, RRT*, and APEI-RRT* planners for TurtleBot3 navigation. RRT* reduced path length by 8.5% and path nodes by ~75% vs RRT.
ROS2GazeboPython
TurtleBot4 Vision-Based Navigation
Real-time perception pipeline for navigation and object detection — arrow-sign following, UMD logo detection, dynamic object TTC estimation.
OpenCVRVizTurtleBot4
Structure from Motion: 3D Reconstruction
SfM pipeline estimating camera poses and sparse 3D geometry from monocular sequences using fundamental/essential matrices and RANSAC.
Computer VisionRANSACNumPy
3D Gaussian Splatting
Differentiable 3D Gaussian Splatting pipeline for novel view synthesis. Optimized position, covariance, opacity, and color using photometric loss.
PyTorch3DGSRendering
Lane Detection & Road Curvature Estimation
Lane detection pipeline using Hough transforms, bird's-eye-view homography, and polynomial fitting for curvature estimation on driving videos.
OpenCVPython
Differential Drive A* for TurtleBot3
A* path planning with non-holonomic constraints and RPM-based action primitives. Trajectory simulated in Gazebo via ROS2 velocity commands.
ROS2A*Gazebo
MoveIt Pick-and-Place — Franka Panda
Franka Panda motion planning in ROS2 Humble using MoveIt. Collision-aware motions from home → pick → drop → home.
MoveItROS2RViz2
Library Robot — 6-DOF Mobile Manipulator
Simulated book retrieval pipeline in ROS2/Gazebo with autonomous navigation, barcode-based identification, and DH-based manipulation.
ROS2GazeboKinematics
MicroMouse Maze Navigation System
C++17 MicroMouse system for 16×16 maze exploration using DFS, backtracking, conditional replanning, and ROS2 action interfaces.
C++17ROS2DFS
Robust Control via Time Delay Control
Simulation comparing TDC vs PD control for manipulators under gravity, friction, and inertia uncertainty. TDC demonstrated improved tracking.
Control TheoryPython
Panorama Stitching & Homography Estimation
Custom and OpenCV-based stitching pipelines using RANSAC, homography estimation, image warping, and feathering-based blending.
OpenCVRANSAC
SUV-Trailer Robot: CAD, URDF & ROS2
Multi-body SUV-trailer designed in SolidWorks, converted to ROS2 Humble with steering/wheel controllers, LiDAR, and Gazebo teleoperation.
SolidWorksURDFROS2
06

Technical Skills

Programming
PythonC++JavaSQLJavaScript
Robotics & Planning
ROS2GazeboRViz2MoveItRRT*A*
Robot Platforms
UR5e (Real + Sim)Franka Panda (Sim)TurtleBot3/4MuJoCoIsaac SimGazebo
Robot Learning & ML
PyTorchDiffusion PolicyImitation LearningMultimodal ML
Computer Vision
OpenCVRANSAC3D ReconstructionGaussian Splatting
Hardware & Tools
GelSight MiniLiDARUR5eUR3eNVIDIA JetsonRaspberry PiDockerLinuxGit