SRS
Portrait of Siddharth Rama Sushil

Siddharth Rama Sushil

Robotics | Perception | Reinforcement Learning

I design production-ready perception systems and robust sim pipelines that transfer well to hardware — specializing in SLAM, visual–inertial navigation, 6-DoF pose estimation and sim-to-real reinforcement learning.

SLAM Reinforcement Learning Navigation
Get In Touch View Projects Download CV

About Me

Currently pursuing M.S. ECE (Intelligent Systems, Robotics and Control) at the University of California, San Diego. I work on visual–inertial SLAM, 6-DoF pose estimation, semantic segmentation and sim-to-real reinforcement learning used on real robots.

I focus on robust, maintainable systems — combining geometry-first classical methods with learning where it helps. I care about simulation fidelity, domain randomization, and deployment constraints when designing perception and control stacks.

Computer Vision ROS 2 C++ / PyTorch Sim-to-Real
Get In Touch Download CV

Experience

Addverb logo

Addverb Technologies

Robotics Engineer • Sept 2023 – Jul 2025

Lead Architect — Object Recognition & 6DOF Pose Estimation

Designed and deployed object recognition and 6-DoF pose estimation modules for production robotic systems.

  • Made an end-to-end 6-DoF pipeline (detection → pose refinement → verification) ready for deployment.
  • Optimized inference for edge hardware and reduced end-to-end latency for closed-loop control.
  • Built test harnesses, unit tests and validation suites for deterministic QA and regression checks.
  • Integrated pose outputs with robot control stack and safety interlocks for reliable pick/place operations.
6DOF pose estimation demo

Visual–Inertial SLAM

Built a visual–inertial SLAM pipeline fusing IMU + vision to improve localization robustness.

  • Fused IMU data with visual features to maintain tracking through low-texture and dynamic scenes.
  • Implemented bias estimation & online calibration to reduce drift over long runs.
  • Designed failure-recovery strategies (relocalization / loop detection) to minimise downtime in production.
Visual-inertial SLAM demo

Semantic Segmentation for SKU Classification

Applied semantic segmentation to classify SKUs and improve robotic perception in cluttered scenes.

  • Collected and annotated domain-specific datasets and implemented robust augmentation strategies.
  • Trained and deployed compact segmentation models, with quantization for edge inference.
  • Built pipelines to convert segmentation outputs into object masks for downstream pose estimation.
Semantic segmentation demo

Enhanced Simulation Fidelity

Improved sim-to-real transfer by increasing simulator fidelity and variability.

  • Added realistic sensor/actuator noise models, IMU bias simulation and jitter to the sim stack.
  • Implemented structured domain randomization (lighting, textures, physics) to cover real-world variance.
  • Simulated network latency and packet loss to validate robustness of distributed components.
Simulation demo

On-site Deployment & Integration

Led on-site deployments ensuring smooth integration into customer environments.

  • Coordinated integrations with partners (Siemens, ISITEC, Adidas) and validated system performance in-situ.
  • Performed live debugging, root-cause analysis and produced runbooks for field teams.
  • Delivered deployment documentation and operator training for production handover.
On-site deployment demo
DRDO logo

Defence Research and Development Organisation (DRDO)

Research Intern — Counter-measure Division • Jan 2023 – Sept 2023

Conducted research on CNNs for infrared small target detection, developed preprocessing pipelines for low SNR IR images, and applied deep-learning feature aggregation to make models robust to practical field noise patterns.

  • Designed preprocessing (denoising / contrast enhancement) to improve target SNR in IR frames.
  • Built pixel-level annotated datasets and evaluation protocols for small-target detection under noise.
  • Explored feature-aggregation techniques to boost performance in low-SNR conditions and wrote technical reports summarizing results.
  • Prototyped models and validated on edge-capable hardware for real-time inference feasibility.
DRDO IRSTD

Featured Projects

SLAM System

Robust Plane Segmentation for 3D Pose Estimation

Geometry-first plane segmentation module that runs in real-time to support high precision pose estimation pipelines.

C++ Bundle Adjustment Sensor Fusion
Crane Hook Pose Tracker

6-DoF Pose Tracker for Crane Hooks

Plug-and-play smart box for tracking crane hook position & orientation (hackathon project with L&T).

Unity C++
Sewer Monitoring Robot

Sewer Monitoring Robot

Underground pipeline mapping & structural analysis robot for efficient inspection and maintenance planning.

Unity C#
UV Disinfectant Robot

UV Disinfectant Robot

Autonomous UV-disinfection robot with optimized zap patterns to maximize area coverage while saving power.

C++ Haptic Rendering
Ground Plane Segmentation

Efficient RANSAC Ground Plane Segmentation

IMU-aided RANSAC algorithm that reduces computation and improves real-time performance for ground plane detection.

C++ OpenCV
Custom A* Path Planner

Custom Path Planning Algorithm

Lightweight ROS-independent A* planner with smoothing designed for resource-constrained boards.

C++ Path Smoothing

Achievements

Gold medal ceremony

Gold Medal B.Tech — 2024

Awarded the D&M Proficiency Gold Medal for best project across the entire B.Tech batch of 2023.

 TIGC prototype demo

Top 10 TIGC — 2024

Ranked in the top 10 at TIGC hackathon (La Trobe University), awarded AUD 1,000 in prototype funding.

 Techgium contest moment

AIR-1 Techgium — 2022

Secured AIR-1 among 800+ participants in an L&T hackathon, winning prize for crane-safety solution.

Personal Blog

I write blogs to share lessons, mistakes, and breakthroughs. My goal is to make learning smoother for aspiring roboticists and inspire them to push through the same challenges I once faced.

Article image 1

Hands on SLAM Frontend: Landmark Lifecycle

In Part 1, I walked through my early experiments with feature detection, matching, and triangulation. I asked a question: if triangulation works so well, why not just keep doing it forever between consecutive frames? ...

Read on Medium →
Article image 2

Hands-On SLAM: My Experience Building a Visual Frontend Pipeline

You’re staring at a research paper about SLAM, and there’s this intimidating diagram with boxes labelled “Bundle Adjustment,” “Loop Closure,” “Pose Graph Optimisation.” Your brain immediately goes into overwhelm mode. Where do you even start? ...

Read on Medium →
Article image 3

Demystifying SVD: The Math Powering Visual SLAM

Singular Value Decomposition sounds intimidating, but it’s actually one of the most powerful and practical tools that helps robots understand what they’re seeing. Think of it as a mathematical Swiss Army knife that solves many of the core problems in robot vision...

Read on Medium →
Article image 4

Epipolar Geometry in Practice: Build, Visualize, Understand

I built a tiny C++/OpenCV repo you can run locally to feel epipolar geometry instead of just reading about it...

Read on Medium →
Article image 5

Foundations of Robot Perception: A Beginner’s Guide to the Math

When you’re building a robot that needs to see the world, you’re basically asking it to solve puzzles like: “Where is that object in 3D space based on what I see in this 2D camera image?” ...“How did the robot move between these two camera shots?”

Read on Medium →
Article image 6

Building a Robust Plane Segmentation Pipeline for 3D Pose Estimation in Robotics

When working on industrial automation systems, flashy CV models and deep learning pipelines often get the spotlight. But sometimes, it’s the quieter, lower-level problems — like reliably extracting a plane from a noisy point cloud — that make or break a robotic system...

Read on Medium →

Get In Touch

Contact Information

Feel free to reach out if you have a question, or just want to connect.

sramasushil@ucsd.edu
+1 (858) 214-6356
San Diego, CA
Or Email Directly