Embodied perception & mapping
Multimodal scene understanding for navigation and manipulation: geometry-aware representations for localization, mapping, and decision-making under real sensing constraints.
About
I’m a robotics and computer vision researcher working on visual mapping and localization, 3D reconstruction, and multimodal scene understanding for embodied systems. My aim is to give robots representations of the world that are geometrically faithful, semantically meaningful, and robust enough to survive deployment.
I completed my PhD at the University of Adelaide under Prof. Ian Reid at the Australian Centre for Robotic Vision, on real-time, structure-aware, object-centric semantic visual SLAM — integrating geometric estimation with object- and scene-level priors for reliable operation in complex environments. Between my PhD and my current role I spent several years in industry, deploying real-time SLAM, mapping, and situational-awareness systems on field and warehouse robots. That experience still shapes how I work: runtime, sensing limits, and failure modes are first-class concerns, not afterthoughts.
Research directions
3D reconstruction & neural representations
Learning-based reconstruction and neural rendering — Gaussian splatting and NeRF — with an emphasis on geometric consistency and generalization beyond the training distribution.
Vision-language reasoning for robots
Using VLMs over structured visual evidence for robot introspection: detecting and localizing failures in long-horizon tasks, producing grounded explanations, and proposing corrections.
Interests
- Visual SLAM
- 3D Reconstruction
- Gaussian Splatting & NeRF
- Geometric Foundation Models
- Vision-Language Models
- BEV Scene Understanding
- Multi-Sensor Fusion
- Autonomous Navigation
- Factor-Graph Optimization
News
- Jan 2026 KITE accepted at ICRA 2026 — keyframe-indexed tokenized evidence for VLM-based robot failure analysis.
- Dec 2025 G³Splat preprint released — geometrically consistent, generalizable Gaussian splatting on VGGT and DUSt3R backbones. Currently under review.
- Jan 2025 TANGO accepted at ICRA 2025 — traversability-aware navigation with local metric control for topological goals.
- Oct 2024 Received the Outstanding Reviewer Award at ECCV 2024, Milan, Italy.
- Jun 2024 BEVPose accepted at IROS 2024 — pose-guided multi-modal BEV alignment for scene semantics.
- Jan 2024 RoboHop accepted at ICRA 2024 — segment-based topological maps for open-world visual navigation.
Publications
Selected work, newest first. A complete list is available on Google Scholar.
Under review
G³Splat: Geometrically Consistent Generalizable Gaussian Splatting
Injects surface-normal and depth priors into pose-free generalizable Gaussian splatting built on geometric foundation models, improving reconstruction geometry in textureless and complex indoor scenes.
ICRA 2026
KITE: Keyframe-Indexed Tokenized Evidence for VLM-Based Robot Failure Analysis
A training-free front-end that compresses long robot-execution videos into compact, interpretable evidence, substantially improving VLM failure detection, identification, and localization.
ICRA 2025
TANGO: Traversability-Aware Navigation with Local Metric Control for Topological Goals
An RGB-only topometric navigation pipeline combining global topological planning with local metric trajectory control for zero-shot, long-horizon navigation without 3D maps.
IROS 2024
BEVPose: Unveiling Scene Semantics through Pose-Guided Multi-Modal BEV Alignment
Uses sensor pose as a supervisory signal to align camera and LiDAR bird’s-eye-view embeddings, outperforming fully-supervised BEV map segmentation with a fraction of the annotated data.
ICRA 2024
RoboHop: Segment-based Topological Map Representation for Open-World Visual Navigation
A segment-level topological representation of the environment that supports zero-shot navigation driven by open-vocabulary language queries.
ICRA 2019
Real-Time Monocular Object-Model Aware Sparse SLAM
Integrates deep-learned object detections and CNN shape priors as quadrics into a monocular SLAM factor graph, refining object geometry while retaining real-time performance.
ACCV 2018
Structure Aware SLAM using Quadrics and Planes
Represents generic objects as quadrics and scene structure as planes in a unified monocular SLAM map, with shape and plane–plane constraints tying semantics to geometry.
Experience
-
2022 – present
Postdoctoral Research Fellow
Australian Institute for Machine Learning (AIML), The University of Adelaide
- Generalizable 3D reconstruction & novel-view synthesis — geometrically consistent Gaussian splatting on VGGT and DUSt3R backbones, folding surface normals and depth into rendering-based reconstruction (G³Splat).
- TAGVIEW (defence collaboration) — stereo motion estimation for off-road navigation in GPS-denied environments, fusing LiDAR, camera, GNSS/INS, and satellite modalities into BEV maps.
- Multi-modal transformer perception — data-efficient ground-plane representations using cross-attention for sensor fusion (BEVPose).
- Topometric mapping & language-guided navigation — segment-level topological representations enabling zero-shot, open-vocabulary navigation (RoboHop, TANGO).
- RECAP (with MITRE Australia) — grounded failure detection, explanation, and correction for long-horizon tasks, augmenting VLA backbones with reasoning and memory layers.
- Leadership & mentoring — supervising Master’s students and leading interdisciplinary experiments across robotics and computer vision.
-
2020 – 2022
Senior Computer Vision & AI Research Engineer
- Built real-time SLAM, mapping, and situational-awareness systems for the Autonomous Load Handling System (ALHS), automating loading and unloading of ISO containers.
- Developed perception modules for payload detection and localization, enabling safe navigation and arm control in outdoor operational settings.
- Led cross-functional software integration, holding real-time performance and reliability under varied lighting and terrain.
-
2019 – 2020
Computer Vision & SLAM Research Engineer
KikTech Robotics — startup environment
- Designed and deployed robust SLAM and mapping frameworks for autonomous warehouse vehicles operating around the clock in highly dynamic logistics settings.
- Adapted vision pipelines to severe illumination change and dynamic obstacles, reducing localization drift and operational downtime.
- Balanced rapid prototyping against commercial constraints, integrating scheduling and path-planning modules for continuous operation.
-
2016 – 2019
PhD Researcher
Australian Centre for Robotic Vision, The University of Adelaide
- Developed real-time object-based semantic SLAM with novel plane and quadric representations driven by CNN detections.
- Demonstrated real-time performance through specialised factor-graph formulations integrating shape and plane–plane constraints into a unified map.
- Collaborated across labs on the ACRV Scene Understanding project, bridging geometric reconstruction and semantic interpretation.
Teaching & Supervision
Ten years of university teaching across programming, algorithms, and AI — from first-year foundations to postgraduate coursework and research supervision.
2024 – 2026
IT Lecturer & Subject Coordinator
Kaplan Business School
Owned curriculum, assessment, and delivery for postgraduate IT subjects; handled grade moderation and academic liaison.
- Advanced Programming
- Algorithms & Data Structures
- AI & ML in IT
2022
Course Coordinator & Lecturer
The University of Adelaide
Coordinated Algorithm & Data Structure Analysis (Semester 2, 2022), delivering lectures and managing the tutoring team.
2017 – 2026
Lecturer — Degree Transfer & Pre-Master Programs
The University of Adelaide College & Kaplan International College
Developed and delivered modules across computing and electronics; recognised in 2018 for teaching excellence on the basis of SELT results.
- Object-Oriented Programming (C++)
- Foundations of Computer Science (Java)
- AI Technologies
- Computer Networks & Applications
- Programming MATLAB/C
- Digital Electronics
- Electronic Systems
- Scientific Computing
Postgraduate supervision
Robotic Navigation using LLMs in Dynamic Environments
M. T. Aamir · Master’s project, 2024
Integrated large language models with 2D LiDAR, YOLO-based detection, and NAV2 for spatial reasoning in dynamic indoor navigation.
3D Gesture Recognition for Human-Robot Interaction
X. Hu · Master’s project, 2024
Built a gesture recognition pipeline using deep learning for real-time 3D motion tracking and feedback across varied environments.
RoboCup — Omid Robotics Team
Long before SLAM, robot soccer taught me how real-time autonomy actually behaves. In 2008 I co-founded the AI group of the Omid Robotics Team, competing in the Small Size League of RoboCup — a hybrid centralised/distributed multi-agent problem where perception, planning, and control all run inside a few milliseconds.
What I built
- A torque-based controller for the robots’ brushless DC motors, giving better tracking response and robustness across playing surfaces than a conventional velocity controller.
- High-level strategy and RRT-based path planning with time-optimal trajectory generation over the planned path.
- A 3D simulator and monitoring GUI for multi-robot soccer, used to prototype and test strategies before matches.
Competition results
- 2010 Round of 8, International RoboCup, Singapore
- 2010 2nd place, RoboCup IranOpen, Tehran
- 2010 3rd place, Khwarizmi & AUTCUP RoboCup, Tehran
- 2009 Qualified, International RoboCup, Graz, Austria
Talks & Service
Invited talks
-
Semantics and Objects in SLAM
Professional service
Computer vision & machine learning Reviewer for CVPR, NeurIPS, ICCV, ECCV, and BMVC since 2023 — Outstanding Reviewer, ECCV 2024.
Robotics Reviewer for ICRA, IROS, IEEE Transactions on Robotics (T-RO), and IEEE Robotics and Automation Letters (RA-L) since 2019.