Spatial Intelligence & Agentic AI
From Persistent Worlds
to Self-Evolving Agents.
Building agents that understand 3D worlds, act through world models, and improve through experience.
Hao Wang is a tenure-track Assistant Professor in the AI Thrust at HKUST(GZ). He received his Ph.D. from Nanyang Technological University, Singapore. His research lies at the intersection of spatial intelligence and agentic AI, with more than 70 academic paper published. He serves as an Area Chair for ACL ARR and a Senior Program Committee (SPC) member for AAAI.
Latest updates.
Persistent Worlds
Using RGB input only, HorizonStream delivers stable streaming 3D reconstruction beyond 10K frames, with constant memory and linear time.
Actionable World Models
Moving world models from plausible generation to controllable prediction and measurable decision utility.
AI that Improves AI
A unified view of AI systems that improve data, training, evaluation, workflows, and eventually themselves.
Selected highlights.
01 Spatial IntelligenceLong-horizon 3D perception, reconstruction, SLAM, and human-centered spatial intelligence
HorizonStream
RGB-only streaming 3D reconstruction beyond 10K frames.
KiloGS-SLAM
Monocular 3D Gaussian SLAM for kilometer-scale outdoor scenes.
LongStream
Streaming autoregressive visual geometry for long sequences.
VGGT4D
Training-free 4D reconstruction by mining motion cues from visual geometry transformers.
MultiGO++
Geometry-texture collaboration for monocular clothed-human reconstruction.
MotionGRPO
RL post-training for diffusion-based egocentric motion recovery.
FastAnimate
Learnable template construction and pose deformation for fast 3D avatar animation.
S3PO-GS
Scale-consistent RGB-only Gaussian SLAM for outdoor scenes.
RegGS
Unposed sparse-view Gaussian splatting through 3DGS registration.
OpenGS-SLAM
RGB-only Gaussian splatting SLAM for unbounded outdoor scenes.
GraphGS
Graph-guided reconstruction of large open scenes from images.
GVKF
Efficient open-scene surface reconstruction with Gaussian voxel kernels.
02 World ModelsFrom predictive environments to actionable and self-improving agents
ReCAPA
Hierarchical predictive correction that prevents cascading failures in embodied agents.
EmbodiedWM
A capability framework for plausible, controllable, and actionable embodied world models.
EvolveNav
Self-evolving rule memory and outcome-aware reasoning for zero-shot navigation.
AI4AI
A unified view of AI systems that improve data, training, evaluation, workflows, and themselves.
03 Game AIMultimodal, strategic, and socially intelligent agents
CaM-Wolf
Causal-aware multimodal agents for social deduction games.
The Stackelberg Speaker
Strategic persuasive communication for social deduction agents.
VistaWise
A cost-effective Minecraft agent grounded by a cross-modal knowledge graph.
CausalMACE
Causality-empowered multi-agent collaboration in Minecraft.
MultiMind
Multimodal reasoning and theory of mind for Werewolf agents.
LLM-Based Agent Society
Studying collaboration, confrontation, and social behavior among LLM agents in Avalon.
Build intelligent worlds together.
Open to joint research and industry collaboration in 3D spatial computing, digital twins, embodied and game agents, generative AI, and self-evolving systems.