Personal Homepage
My current work centers on policy learning for embodied AI, dual-arm and humanoid VLA systems, and task planning and execution frameworks for robots. I have also worked on intelligent transportation systems, including multi-agent trajectory prediction, traffic flow modeling, and co-simulation.
About
My research interests include embodied AI, vision-language-action models, latent world models, multi-agent systems, and agent operating systems for robots. In my current work, I participate in building an agent operating system for robots by organizing language understanding, navigation, action generation, and speech interaction into reusable skills, and deploying these capabilities on platforms such as Unitree G1. I also study VLA frameworks with latent world models to improve modeling of future state evolution.
In intelligent transportation systems, I have also worked on multi-agent trajectory prediction, traffic flow and occupancy prediction, lane-level road network extraction from remote sensing imagery, and energy evaluation based on co-simulation and vehicle dynamics modeling.
Focus
I work on policy learning, VLA reproduction, action generation, and real-world deployment for dual-arm and humanoid robots, with emphasis on the coordination of language interaction, perception, and execution at the system level.
I explore integrating latent world models into robot manipulation so that future state evolution can be modeled from heterogeneous multi-source data, improving policy generalization, temporal modeling, and physical consistency.
I study trajectory prediction, traffic flow modeling, road network extraction, and co-simulation in complex traffic scenarios, integrating multi-agent interactions, map priors, and spatiotemporal dependencies into interpretable learning frameworks.
Timeline
Expected to start in Aug. 2026. Research interests include embodied AI, dual-arm and humanoid VLA algorithms, and embodied multi-agent systems.
Conducted research on intelligent transportation systems, autonomous driving, and co-simulation under the supervision of Prof. Yunli Shao.
Focused on spatiotemporal data mining and traffic prediction.
Designed a collaborative "big-brain/small-brain" policy system for embodied agents based on Aggno and the Qwen family; abstracted capabilities such as TTS, VLA, and navigation into reusable skills; and built a Pipecat-based online speech interaction agent to close the loop from natural language input to robot execution.
Worked on AIS data cleaning, trajectory reconstruction, and ship behavior recognition, and extracted key maritime nodes using trajectory clustering and spatial statistics.
Built workflows for multimodal time-series data cleaning and causal discovery, and reproduced and evaluated algorithms such as PC, GES, and NOTEARS for dependency analysis in complex systems.
Selected Work
I investigate how latent world models can be integrated into robot manipulation pipelines to address the lack of temporal modeling and physical priors in existing VLA systems. The goal is to improve policy generalization and future-state reasoning with heterogeneous data from multiple robot platforms.
Proposed a lane-level road network extraction framework for transportation digital twins, together with a heuristic intersection topology inference algorithm that supports export to standard formats such as OpenDRIVE and SUMO.
Guo, H., Shao, Y., Saroj, A., Xu, G., Yuan, J., Luo, X., Wang, C. (2026). Automatic lane-level road network extraction from aerial imagery for transportation digital twins. ICTD.
Built an XIL co-simulation framework integrating SUMO, a connected vehicle controller, and a Simulink vehicle dynamics model, and analyzed how model granularity and penetration rate affect system-level energy evaluation.
Yuan, J., Guo, H., Shao, Y., Xu, G., Saroj, A., Wang, C. (2025). Impact of vehicle dynamics on the energy evaluation of connected and automated vehicles via anything-in-the-loop co-simulation. TRB Annual Meeting.
Proposed a traffic flow-aware occupancy prediction framework for connected and automated vehicles, explicitly modeling spatiotemporal coupling between adjacent road segments to improve long-horizon consistency.
Guo, H., Shao, Y. (2025). Traffic flow aware occupancy prediction for energy and mobility centric connected and automated vehicles. MECC.
Built an end-to-end marine traffic flow forecasting framework that decomposes traffic flow into trend and periodic components, jointly modeling spatiotemporal and static-dynamic dependencies for both accuracy and interpretability.
Yuan, Y., Guo, H., Fan, Z., Peng, Y., Zhang, J., Song, X., Shibasaki, R. (2024). A decomposable multi-fusion spatiotemporal marine traffic flow forecasting algorithm. Transportation Research Part E.
Proposed HHGNN to model high-order group interactions in complex traffic scenes through heterogeneous hypergraphs, with type-aware message passing and local high-definition map information.
Guo, H., Peng, Y., Fan, Z., Zhu, H., Song, X. (2024). HHGNN: Heterogeneous hypergraph neural network for traffic agents trajectory prediction in complex scenarios. IEEE ICRA.
Capabilities
Contact
If you are interested in embodied AI, robotics, VLA systems, intelligent transportation systems, or interdisciplinary research collaboration, feel free to contact me by email.