Hello
I am Linfeng Zhao (赵林风), a Postdoctoral Scholar at Stanford University, working with Prof. Mykel Kochenderfer, Prof. Jeannette Bohg, SISL, IPRL, and the Stanford Robotics Center.
I finished Ph.D. (2025) at Khoury College of Computer Sciences of Northeastern University, advised by Prof. Lawson L.S. Wong, where I closely collaborated with MIT LIS group and Prof. Leslie Kaelbling, and Prof. Robin Walters. I interned at Meta, Boston Dynamics AI Institute, Amazon, and Microsoft Research Asia. Before, I worked with Prof. Hao Su at UC San Diego (2018-19).
My research focuses on building human-level general-purpose agents that can act in the physical world—robots that navigate homes, manipulate objects, and accomplish long-horizon tasks in open-world scenarios with unseen environments and goals. I develop abstractions for decision-making that decompose complex behaviors into compositional building blocks. I develop learning and planning approaches to enable agents to reason about the world and plan their actions at decision time for scalable, generalizable, and efficient decision-making systems.
Learning and Planning with Structured Abstraction for Embodied Decision-Making
PDFResearch highlight
Retriever
We develop a programming model for closed-loop robot agents that perceive, reason, and act in parallel.
Updates
Gave a series of talks on building composable robot systems for long-horizon manipulation. I highlighted Retriever, our programming model for closed-loop robot agents, alongside my work on robot planning and manipulation. Hosts included Carnegie Mellon University (Changliu Liu's group, AirLab, and Andrea Bajcsy's Intent Lab), Stanford (Shuran Song's group), and Nanyang Technological University (Yoonchang Sung's group).
Presented Retriever: Closing the Perception-Reasoning-Action Loop for Long-Horizon Robot Manipulation at the Saturday Robotics World Models Reading Club (recording) and Sundae Robotics.
Open-sourced Retriever, a programming model and runtime for composing asynchronous robot programs (paper, code, docs).
TL;DR
Agentic behavior takes more than a model call. Retriever gives perception, memory, planning, and action a shared programming model without forcing them onto one clock. In the drawer search, the robot remembers where it found the pepper and goes straight there the next time. See the demo.
Our GEM paper, Equivariant Open-vocabulary Pick and Place via Language Kernels and Patch-level Semantic Maps, was selected as a RA-L 2025 Best Paper Award Honorable Mention.
Relocate to California for Postdoc research at Stanford University.
Robot Demos
Research Focus
Abstraction and Representation
Learning structured representations (symmetry, compositionality) for efficient generalization.
Planning and World Modeling
Developing world models and planning algorithms for long-horizon reasoning.
Test-time Scaling
Scaling decision quality with more compute and interaction.
Mobile Manipulation
Enabling diverse manipulation skills on mobile robot platforms.
Selected Publications
MessyMem: Learning-from-Doing Memory for Mobile Manipulation
TL;DR
Action Map Policy: Learning 3D Closed-loop Manipulation via Pixel Classification
TL;DR
Pix2Act: Image-Space Manipulation Policies with Equivariant Augmentation
TL;DR
Retriever: Composing the Perception-Reasoning-Action Loop for Long-Horizon Manipulation
TL;DR
Learning and Planning with Structured Abstraction for Embodied Decision-Making
TL;DR
Seeing is Believing: Planning to Perceive with Foundation Models and Act Under Unknowns
TL;DR
Practice Makes Perfect: Planning to Learn Skill Parameter Policies
TL;DR
Equivariant Open-vocabulary Pick and Place via Language Kernels and Patch-level Semantic Maps
TL;DR
E(2)-Equivariant Graph Planning for Navigation
TL;DR
Integrating Symmetry into Differentiable Planning with Steerable Convolutions
TL;DR
Scaling up and Stabilizing Differentiable Planning with Implicit Differentiation
TL;DR
Toward Compositional Generalization in Object‑Oriented World Modeling
TL;DR
