Postdoctoral Research Associate
Australian Artificial Intelligence Institute
University of Technology Sydney
Email: firstname.lastname at uts.edu.au
I am a Postdoctoral Research Associate at the University of Technology Sydney (UTS), working with Prof. Ling Chen. I received my Ph.D. from The University of Queensland, supervised by Prof. Xin Yu.
My research focuses on agentic AI and multimodal learning. I am interested in how AI systems connect perception with reliable action: understanding sounds and visual scenes, grounding instructions in interfaces, and carrying out useful tasks in real-world environments.
My current work explores computer-use agents for long-horizon workflows and the diagnostic evaluation of multimodal foundation models. Previously, I developed object-centric audio-visual alignment methods for sounding-source segmentation. I have also worked with CSIRO Data61, NetEase Fuxi AI Lab and Alibaba Tongyi Lab.
Selected work across agentic AI, multimodal understanding, audio-visual learning, 3D/4D generation, digital humans and agriculture engineering. * Equal contribution.
SATEBench evaluates human-centric spatial understanding through active exocentric exploration, contrasting multi-turn view selection with passive multi-view observation.
An open stack for generalist computer-use agents that combine visual interaction with programmatic tools.
Reusing intermediate visual evidence within a single forward pass for accurate, efficient GUI grounding.
ATMem turns passive interaction records into active task-driving memory for long-horizon GUI agents.
A benchmark for mobile agents spanning cross-app workflows, user interaction and MCP tools.
Treating diverse instructions as reasoning pathways for more reliable GUI grounding.
A two-stage approach first learns from diverse instructions, then uses reinforcement learning to select and combine useful perspectives. UI-Ins-7B serves as the executor in an AndroidWorld agent with a reported 74.1% success rate.
Uni4R brings 4D reconstruction and point tracking together through continuous velocity fields.
Boundary-aware features and adaptive warping improve mesh reconstruction from sparse views.
Generating coherent and diverse co-speech gestures from in-the-wild speech with a large-scale gesture dataset.
Evaluating human understanding through questions that require complementary evidence from multiple camera views.
Distinguishing robustness to redundant views from the ability to combine evidence across views.
Foundation GUI agents integrating user interaction, MCP tools, device–cloud collaboration and online reinforcement learning.
Deriving distinct sound semantics and removing irrelevant audio for precise visual grounding.
DDESeg addresses overlapping sound sources and variation in the sounds produced by the same object. Semantic derivation enriches source-specific representations, while dynamic elimination removes audio information that does not match visible sounding regions.
Local audio-visual alignment and uncertainty estimation for changing sounding objects.
Audio-guided interactions group visual features by their response to sound. An uncertainty module accounts for spatial and temporal ambiguity when objects frequently switch between sounding and silent states.
Separating dynamic and static frame features for video-to-4D generation.
A benchmark for grounding sounding objects across long videos, including silent intervals.
LU-AVS studies where and when objects produce sound in untrimmed videos. Frame-level temporal annotations, segmentation masks and bounding boxes support evaluation beyond short, continuously sounding clips.
Learning an expression representation to transfer subtle facial motion across 3D avatars.
FreeAvatar combines an expression foundation model with a multi-avatar animation model. Learned expression features provide perceptual supervision, helping capture fine-grained expressions without relying on geometric constraints.
Combining visual masks and audio semantics to suppress background noise and off-screen sounds.
A two-stage framework first identifies potential sounding objects and audio categories. An audio-visual semantic integration strategy then associates these predictions through a hierarchical audio-visual tree.
Modeling emotion and rhythmic beats in speech to generate expressive and diverse 3D gestures.
Interactive hairstyle editing using parametric strokes to control local hair shape and appearance.
Learning audio-visual semantic correspondence to distinguish sounding objects from silent distractors.
A silent-object-aware objective reduces visual saliency bias. Audio category scores are associated with potential object masks to establish cross-modal semantic correspondence.
Predicting diverse 3D hand gestures from body motion while modeling the two hands separately.
An energy-saving adder neural network for cucumber disease classification.
Efficient cucumber disease detection in natural scenes with complex backgrounds.
Postdoctoral Research Associate · Australian Artificial Intelligence Institute
Ph.D. · Object-Centric Audio-Visual Alignment for Sounding Source Segmentation
Master’s degree
