Hu Tianrun

Hu Tianrun 胡天润

PhD Student · School of Computing @ NUS

"wait, this actually works?"

I work on mobile manipulation robots that reason, look, and act in unknown and changing environments.

Email Scholar GitHub ORCID LinkedIn

About Me

I am a PhD student at SoC, NUS, under the supervision of Prof. Harold Soh. I have been awarded the A*STAR Graduate Scholarship (Computing). Previously, I was a Research Engineer at SSI, under the supervision of Prof. David Hsu. During that time, I worked closely with Dr. Hanbo Zhang. Previously, I was an undergraduate in Computer Engineering at CCDS, NTU under the guidance of Prof. Lam Siew Kei.

Research Interests

My research focuses on real-world mobile manipulation, with an emphasis on reasoning in unknown and dynamic environments. I develop representations and planning methods that enable robots to reason under partial knowledge, direct sensing toward task-relevant uncertainty, and adapt their actions as the environment changes.

Joint Reasoning over Perception and Action › Can we reason about where to look and how to act under one unified objective, instead of treating perception and action as two separate pipelines?
[Visibility-Aware Mobile Grasping] RSS 2026 Workshop (ExWBC), Oral Presentation. Joint gaze and base-motion planning under limited directional sensing to grasp objects in dynamic environments.
[Unified Seeing, Avoiding, Grasping] ongoing. A probabilistic formulation that casts the three operations as expected overlap between future probability fields, with gaze maximizing collision risk to expose it and motion minimizing execution risk.
Scene Understanding for Action › Scene representations and grounding are often evaluated as pure vision tasks and then adapted to action. I want representations that are designed for action from the start, so the loop closes from the perception side as well.
Action-Oriented Object Representations
[MimicFunc] CoRL 2025. Functional correspondence as a perception representation built for tool use, transferring from a single human video.
[Chain-of-Action] NeurIPS 2025. Trajectory autoregressive modeling anchored on key object states, tying the state representation directly to action.
[Object-Centric Policy] ongoing. Contact points and spatial vectors in object coordinates, transferring across embodiments and viewpoints.
Scene Representations and Grounding
[Dynamic Scene Graph] ongoing. Relational scene structure that tracks object states under partial observation to support long-horizon planning.
[Uncertainty Estimation for Visual Grounding in the Open World] in submission. Calibrated uncertainty over open-world grounding so that downstream planning can reason about what the perception system does and does not know.
Reasoning for Complex Real-World Tasks › Extending joint reasoning from single-skill mobile grasping to open-world, multi-step tasks that a robot has to handle in homes and public spaces.
[Hypothesis-driven Model Expansion under Uncertainty] RSS 2026. Reasoning over hypotheses about an open-world environment and expanding the planning model on demand.
[Robi Butler] ICRA 2025. Remote multimodal interaction with a household robot assistant via language, gesture, and demonstration.