Curriculum vitae / Robotics software engineer
Bavantha Udugama
Robotics software engineer with 8 years in robotics across research and industry, starting with award-winning undergraduate research in autonomous ground robots. I take mapping, perception and navigation from algorithm design to deployment on real robots. I built and flew a custom autonomous drone with onboard perception, and led a three-engineer team that deployed more than 30 commercial robots.
- Role
- Robotics software engineer, 3D perception and computer vision
- Based
- Netherlands
- Available
- Immediately
- Focus
- Perception, SLAM and navigation, from algorithm to deployment
At a glance
- 8 yrs
- in roboticsResearch and industry: University of Peradeniya, Synopsys, Premier Genie, Expert Hub Robotics, University of Twente
- 4
- first-author papersISPRS Annals, IROS 2025, ICRA 2026 SRRA, ISPRS Journal under review
- 30+
- robots deployedHumanoid service robots across Expo 2020 Dubai, exhibition stands, healthcare, and reception sites
- 25.53 fps
- on-device perceptionM2H-MX-L, TensorRT FP16, Jetson Orin NX 16GB
Experience
Doctoral researcher, robotics and computer vision
Faculty ITC, University of Twente Enschede, Netherlands
PhD project: autonomous real-time UAV mapping with spatial perception for indoor applications. Supervised by George Vosselman and Francesco Nex.
- Developed C++ software for monocular visual-inertial SLAM and 3D scene-graph mapping in ROS 2 (Mono-Hydra++). Sparse predicted-depth factors in the factor graph cut trajectory error by 61% on ScanNet against monocular GO-SLAM while holding 25 to 30 Hz.
- Designed the multi-task deep learning architecture for joint depth estimation and semantic segmentation (M2H, M2H-MX). M2H appeared at IEEE/RSJ IROS 2025; M2H-MX, accepted at the IEEE ICRA 2026 SRRA Workshop, improves NYUDv2 mIoU by 6.6% and depth RMSE by 9.4% over the best tested baseline.
- Built a custom autonomous drone and deployed onboard perception on its Jetson Orin NX 16GB, exported through ONNX and optimised with TensorRT FP16, processing images at 25.5 frames per second.
- Developed ATLAS, an autonomous exploration algorithm on top of that mapping system: a compact belief graph of places and open ends keeps goal selection real time, trained in Isaac Sim from planner demonstrations and deep reinforcement learning, integrated with Mono-Hydra++ over ROS 2 and deployed on the drone.
- Prepared and taught practical exercises for two master's courses in robot perception and autonomy, and supervised three master's thesis students.
The loop this PhD closes, in four clips.
01
Perception and scene graph
The whole pipeline on ScanNet scene 0000. M2H-MX predicts depth and semantics from the image stream, Mono-Hydra++ consumes those predictions inside the visual-inertial odometry and the mapping backend, and the hierarchical scene graph is built in real time. Playback is 16x. M2H-MX predicts metric depth and semantics in a single pass, and Mono-Hydra++ turns them into a hierarchical 3D scene graph. The same stack, run across different environments.
02
Robustness on device
Depth-backed visual-inertial odometry on the drone's Jetson, walked down and back up the stairs at home. The label switches between MOVING and ZUPT STATIONARY: when the platform is still, zero-velocity updates hold the estimate in place instead of letting it drift. Predicted depth, tracked features, and predicted semantics are the panels on the left. The VIO running on a Jetson Orin with a ZED X One, holding pose down the stairs and back up. This is the robustness test for the algorithm.
03
The drone
Take-off and a pass down the corridor. Everything the mapping runs on is carried onboard. The quadrotor built to deploy this stack, carrying and computing all of it onboard.
04
Exploration
Autonomous exploration in simulation. The robot holds an occupancy map it has built itself and commits to one certified route at a time, so the map fills in region by region rather than by wandering. 60 second excerpt. Still at simulation level, in NVIDIA Isaac Sim and the lab: the map it built decides where to look next. Deployment on the drone comes next.
Senior robotics software engineer and project lead
Expert Hub Robotics Dubai, United Arab Emirates
Joined as the founding robotics engineer and was promoted within 10 months to lead a three-engineer team.
- Developed navigation software and led deployments of more than 30 robots for more than 10 enterprise clients, including a six-month run at Expo 2020 Dubai, resolving sensor and software integration issues on site.
- Integrated ROS and GMapping navigation, LiDAR and 3D obstacle sensing with vendor robot SDKs and voice assistance for autonomous visitor guidance, recovering localisation failures on site in changing crowd conditions.
- Fine-tuned a YOLOv5 face detector for grayscale thermal imagery and deployed it on an Intel NUC with a FLIR camera for contactless temperature screening, running live across enterprise sites.
The deployments in the press
- Have you met Ahmed and Hamda at Expo 2020 Dubai? Gulf News, November 2021The Dubai Police Transport Security Department robots on the Expo 2020 site.
- Pandemic boosts use of AI and robotics: Exness employs robots at Dubai Expo LeapRate, December 2020Three Sanbot Max robots greeting visitors and guiding them to the Exness stand at the Dubai Forex Expo.
The fleet in service, and the map it navigates on.
Robotics engineer, STEM and IoT
Premier Genie Dubai, United Arab Emirates
- Created robotics and IoT prototypes for STEM projects, combining embedded sensors, robot behaviours and Android interfaces.
Corporate applications engineer
Synopsys Inc. Sri Lanka, and Samsung Electronics, South Korea
- Reduced chip verification violations threefold during a two-month on-site assignment at Samsung Electronics, South Korea.
Education
PhD, robotics and computer vision
Faculty ITC, University of Twente Enschede, Netherlands
Expected December 2026
Thesis: Structuring the Seen, Exploring the Unseen. Autonomous real-time UAV mapping with spatial perception for indoor applications.
MSc Robotics Engineering
Middlesex University Dubai United Arab Emirates
Distinction
Thesis: a heterogeneous swarm robot system with distributed data processing at the edge for collective transportation.
BSc Electrical and Electronic Engineering
University of Peradeniya Sri Lanka
Second Class Upper Division, GPA 3.55 of 4.00
Final-year project: a vision and LiDAR mobile robot for reconnaissance in unknown environments, fusing a laser range finder with stereo cameras.
Publications
Four first-author papers from the PhD. Each links to a page with the claim, the numbers, and the code.
- 2026 Mono-Hydra++: Real-Time Monocular Scene Graph Construction with Multi-Task Learning for 3D Indoor Mapping
- 2026 M2H-MX: Multi-Task Semantic and Geometric Perception for Real-Time Monocular 3D Scene Graph Construction
- 2025 M2H: Multi-Task Learning with Efficient Window-Based Cross-Task Attention for Monocular Spatial Perception
- 2023 Mono-Hydra: Real-Time 3D Scene Graph Construction from Monocular Camera Input with IMU
Teaching and supervision
2022 to 2026
Designed lab exercises on YOLO-based human detection and tracking, and on LiDAR-based SLAM, for the Robotics master's courses. Supervised 3 master's thesis students.
Awards
- 2017Best Undergraduate Project
Institution of Engineers Sri Lanka - 2017Best Technical Paper
IET Sri Lanka - 2021MSc awarded with Distinction
Middlesex University Dubai
Skills
Robotics software
Computer vision
Embedded deployment and hardware
Active SLAM and navigation
Timeline
Every position and every milestone in one column, newest first.
-
2026
Mono-Hydra++ under review
ISPRS Journal of Photogrammetry and Remote Sensing
The full system. 0.08 m on the ITC 2nd floor, 0.033 m calibrated ATE on 7-Scenes.
-
2026
M2H-MX accepted
ICRA 2026 SRRA Workshop, Vienna
65.60 mIoU on NYUDv2, and ScanNet average ATE down from 17.59 cm to 6.91 cm inside the mapping loop.
-
2026
Running on the drone
Jetson Orin NX 16GB
M2H-MX-L exported to ONNX and TensorRT FP16 with CUDA Graphs, 25.53 fps and 39.02 ms mean GPU compute at 256x192 on a Jetson Orin NX 16GB. Perception, odometry, mapping, and scene graph construction all run onboard.
-
2025 to present
Learned exploration, in progress
University of Twente
A learned selector over certified routes. Formulation written, experiments not yet run, no results.
-
2025
M2H published
IEEE/RSJ IROS 2025
Four dense tasks in one model. ITC mean error drops to 0.11 m and the loop doubles to 30 fps.
-
2023
Mono-Hydra published
ISPRS Annals
First real-time 3D scene graph from a monocular camera and an IMU. 0.19 m mean error on the ITC 2nd floor at 15 fps.
-
August 2022 to August 2026
PhD candidate, robotics and computer vision
Faculty ITC, University of Twente
Autonomous real-time UAV mapping with spatial perception for indoor applications. Supervised by George Vosselman and Francesco Nex. Thesis: Structuring the Seen, Exploring the Unseen. Expected completion August 2026.
-
October 2020 to June 2022
Senior Robotics Software Engineer and Project Lead
Expert Hub Robotics, Dubai
Founding robotics engineer, promoted within 10 months to lead a team of 3. Built the core ROS navigation stack. Led deployment of more than 30 humanoid robots across Expo 2020 Dubai, healthcare, and reception sites. Fine-tuned and shipped a YOLOv5 thermal face detector running live in a commercial product.
-
October 2019 to September 2020
Robotics Engineer, STEM and IoT
Premier Genie, Dubai
Robotics and IoT prototypes for Dubai school programmes.
-
September 2019 to November 2021
MSc Robotics Engineering
Middlesex University Dubai
Distinction. Thesis on a heterogeneous swarm robot system with distributed data processing at the edge for collective transportation.
-
November 2017 to June 2019
Corporate Applications Engineer
Synopsys Inc., Sri Lanka
Chip verification support, including on-site work at Samsung Electronics in South Korea within the first year.
-
2014 to 2017
BSc Electrical and Electronic Engineering
University of Peradeniya, Sri Lanka
Second Class Upper Division, GPA 3.55 of 4.00. Final-year project: autonomous exploration and mapping for a reconnaissance robot, using a laser range finder fused with stereo cameras. Best Undergraduate Project (IESL) and Best Technical Paper (IET Sri Lanka), 2017.