Embodied AI and Robotic Vision Research Group
We develop embodied AI systems that perceive, reason, navigate, and act reliably in complex open-world environments.
The Embodied AI and Robotic Vision Research Group develops perception, learning, and reasoning capabilities for autonomous robots. The group sits within the Australian Institute for Machine Learning (AIML), Adelaide University, and connects robotics, computer vision, machine learning, and embodied AI.
News
- 2026 — KITE: Keyframe-Indexed Tokenized Evidence for VLM-Based Robot Failure Analysis (Hosseinzadeh, Wong, Dayoub) accepted at ICRA 2026.
- 2026 — SceneEdited: A City-Scale Benchmark for 3D HD Map Updating via Image-Guided Change Detection (Lin, Chin, Garg, Dayoub) accepted at WACV 2026.
- 2025 — Three papers accepted at ICRA 2025: TANGO (traversability-aware navigation), Robust Scene Change Detection with visual foundation models, Effective Tuning Strategies for Generalist Robot Manipulation Policies.
- 2025 — 3D-LLaVA: Towards Generalist 3D LMMs with Omni Superpoint Transformer accepted at CVPR 2025.
- 2025 — Two papers at WACV 2025: To Ask or Not to Ask? (vision-and-language navigation) and Enhancing Embodied Object Detection with Spatial Feature Memory.
- 2025 — Embodied Domain Adaptation for Object Detection accepted at IROS 2025.
Research Themes
Reliable Deep Learning for Robotics
We build perception systems that remain trustworthy beyond closed-set accuracy benchmarks. This includes out-of-distribution and unknown-class detection, open-world object detection, uncertainty estimation for identifying open-set errors, and run-time monitoring of model performance during deployment. The goal is robotic systems that recognise novel objects, flag when inputs violate their training assumptions, and fail safely in unstructured real-world conditions.
Embodied AI and Robotic Vision
We study perception that is active and grounded in an agent acting in the world rather than passive image analysis. This includes embodied object detection with spatial feature memory, embodied domain adaptation for detectors deployed on robots, physically grounded 3D world models for prediction and control, robot manipulation (generalist manipulation policies, data-driven dexterous manipulation), and multimodal perception under degraded or missing sensing.
Open-World Navigation, Mapping, and Localisation
We develop representations and methods that let robots navigate environments they were not explicitly mapped for. This includes topological maps for open-world visual navigation, traversability-aware navigation to topological goals, fine-grained cross-view geo-localisation between ground and satellite imagery, scene change detection with visual foundation models, and vision-and-language navigation that knows when to ask for missing information.
Adaptation, Generalisation, and Foundation Models for Robotics
We work on closing the gap between training and deployment, and on bringing large pretrained models into embodied settings. This includes source-free and online domain adaptation for object detection, predicting and handling class-distribution shift, foundation models for goal-oriented reinforcement learning and exploration, generalist robot manipulation policies and effective strategies for tuning them, and 3D large multimodal models for scene understanding, moving toward vision–language–action systems.
People
.
Group Lead
.
Postdocs / Research Fellows
.
PhD Students
.
Masters / Honours Students
.
Alumni
.
Visitors and Collaborators
.
Join Us
Prospective PhD Students
.
Relevant backgrounds include:
- Robotics
- Computer vision
- Machine learning
- Embodied AI
- Deep learning
- Vision-language-action models
Interested applicants should send:
- CV
- Transcript
- Research statement
- Example paper, thesis, or project
- GitHub or portfolio, where relevant
Postdocs and Visitors
.
Honours / Masters Students
.
Industry Partners
.
Collaborations and Partners
.
Contact
.