Zhihao Lin

Zhihao Lin, Ph.D.

University of Glasgow

Glasgow, U.K.

I received my Ph.D. in Autonomous Systems & Connectivity from the University of Glasgow in June 2026, supervised by Dr. Jianglin Lan. 📬 Open for postdoctoral positions from Sept. 2026.

My research centres on representation learning for reinforcement learning, asking: what should an RL agent learn to see, so that it can act well?

This question grew out of my early work on autonomous driving, where I kept running into the same quiet puzzle: better perception did not automatically lead to better decisions. The gap between seeing and acting never felt like something more data or bigger models would simply close — it seemed to point at something more fundamental about how an agent’s understanding of the world becomes the way it chooses to act. That gap is what I keep returning to.

My Ph.D. work approaches it from three angles:

Geometric policy optimisation (GAC, ICLR 2026): treating bounded action spaces as a geometric constraint rather than an afterthought — replacing Gaussian policies and their ad-hoc squashing with an efficient spherical formulation that decomposes each action into a direction vector and a learnable concentration parameter.
Action Manifold Smoothing (AMS, ICML 2026): stabilising high-dimensional continuous control by replacing point-wise temporal-difference targets with orthogonally-sampled neighbourhood averages, taming the multiplicative Lipschitz-pathway error amplification that makes algorithms like TD3 and SAC collapse.
World-model-guided representation learning (NeurIPS 2026, under review): using a world model not as a simulator but as a structured supervision tool, shaping an encoder whose representations are simultaneously predictive and value-aware.

On the side, I have a deep personal interest in theoretical physics, particularly the thermodynamic and information-theoretic foundations of gravity and cosmology.

I am always happy to chat about RL, world models, embodied intelligence, or the physics of spacetime. Feel free to reach out.

news

Jun 24, 2026	Three papers accepted 🎉 — Dual-Mode SPL-SLAM (co-first author) in IEEE Transactions on Intelligent Transportation Systems, and two first-author papers in IEEE Transactions on Vehicular Technology: Hierarchical Multi-Agent MCTS for Safety-Critical Coordination in Mixed-Autonomy Roundabouts and A Two-Stage Spatiotemporal Trajectory Optimization Framework for Autonomous Lane Changing With Dynamic Risk Fields.
May 15, 2026	My sole-authored paper Action Manifold Smoothing: A Lipschitz Pathway Perspective on High-Dimensional Reinforcement Learning has been accepted to ICML 2026 (CORE A*). 🎉
Dec 15, 2025	My sole-authored paper Beyond Distributions: Geometric Action Control for Continuous Reinforcement Learning has been accepted to ICLR 2026 (CORE A*). 🎉
Dec 10, 2025	Our paper Scalable and Safe Multi-Agent Coordination with Reconstructed Level-k Monte Carlo Tree Search has been accepted to AAMAS 2026 (CORE A*). 🎉
Apr 01, 2025	🏆 Awarded a £1,800 Research Mobility Fund from the University of Glasgow College of Science and Engineering for international research collaboration.

selected publications

ICML
Action Manifold Smoothing: A Lipschitz Pathway Perspective on High-Dimensional Reinforcement Learning

Zhihao Lin

In International Conference on Machine Learning (ICML), 2026

CORE A*

Abs Bib HTML PDF

A Lipschitz pathway perspective on smoothing the action manifold for stable, efficient high-dimensional reinforcement learning.
@inproceedings{lin2026actionmanifold, title = {Action Manifold Smoothing: A Lipschitz Pathway Perspective on High-Dimensional Reinforcement Learning}, author = {Lin, Zhihao}, booktitle = {International Conference on Machine Learning (ICML)}, year = {2026}, note = {CORE A*}, }
ICLR
Beyond Distributions: Geometric Action Control for Continuous Reinforcement Learning

Zhihao Lin

In International Conference on Learning Representations (ICLR), 2026

CORE A*

Abs Bib HTML PDF

A geometry-respecting framework for continuous control that moves beyond unbounded Gaussian policies to honor the intrinsic geometry of bounded action spaces.
@inproceedings{lin2026beyonddistributions, title = {Beyond Distributions: Geometric Action Control for Continuous Reinforcement Learning}, author = {Lin, Zhihao}, booktitle = {International Conference on Learning Representations (ICLR)}, year = {2026}, note = {CORE A*}, }

AAMAS

Scalable and Safe Multi-Agent Coordination with Reconstructed Level-k Monte Carlo Tree Search

Zhihao Lin, Lin Wu, Zhen Tian, and 2 more authors

In International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2026

CORE A*

Bib HTML PDF

@inproceedings{lin2026levelkmcts,
  title = {Scalable and Safe Multi-Agent Coordination with Reconstructed Level-k Monte Carlo Tree Search},
  author = {Lin, Zhihao and Wu, Lin and Tian, Zhen and Lomuscio, Alessio and Lan, Jianglin},
  booktitle = {International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS)},
  year = {2026},
  note = {CORE A*},
}

IEEE TVT

A Two-Stage Spatiotemporal Trajectory Optimization Framework for Autonomous Lane Changing With Dynamic Risk Fields

Zhihao Lin, Zhen Tian, Xianxian Zhao, and 3 more authors

IEEE Transactions on Vehicular Technology, 2026

Bib HTML PDF

@article{lin2026twostage,
  title = {A Two-Stage Spatiotemporal Trajectory Optimization Framework for Autonomous Lane Changing With Dynamic Risk Fields},
  author = {Lin, Zhihao and Tian, Zhen and Zhao, Xianxian and Zhuang, Hanyang and Yang, Ming and Lan, Jianglin},
  journal = {IEEE Transactions on Vehicular Technology},
  year = {2026},
}

IEEE TITS

Contingency-Aware Spatiotemporal Optimization for Safe Autonomous Vehicle Trajectory Planning

Zhihao Lin, Jianglin Lan, Anh-Tu Nguyen, and 1 more author

IEEE Transactions on Intelligent Transportation Systems, 2025

Abs Bib HTML PDF

@article{lin2025contingency,
  title = {Contingency-Aware Spatiotemporal Optimization for Safe Autonomous Vehicle Trajectory Planning},
  author = {Lin, Zhihao and Lan, Jianglin and Nguyen, Anh-Tu and Flynn, David},
  journal = {IEEE Transactions on Intelligent Transportation Systems},
  volume = {26},
  number = {11},
  pages = {18487--18499},
  year = {2025},
}

Pattern Recognit.

SLAM2: Simultaneous Localization and Multimode Mapping for Indoor Dynamic Environments

Zhihao Lin, Qi Zhang, Zhen Tian, and 4 more authors

Pattern Recognition, 2025

Abs Bib HTML PDF

@article{lin2025slam2,
  title = {SLAM2: Simultaneous Localization and Multimode Mapping for Indoor Dynamic Environments},
  author = {Lin, Zhihao and Zhang, Qi and Tian, Zhen and Yu, Peizhuo and Ye, Ziyang and Zhuang, Hanyang and Lan, Jianglin},
  journal = {Pattern Recognition},
  volume = {158},
  pages = {111054},
  year = {2025},
}