Hi, I'm Chan Hee (Luke) Song.송찬희.
I am a researcher at NVIDIA Metropolis, focusing on spatial understanding in vision-language models and their applications in smart spaces, robotics, and other interactive physical environments.
I received my PhD from The Ohio State University, where I was advised by Yu Su.
I have interned at Google Research, Adobe Research, and NVIDIA Research.
I'm always happy to discuss potential collaborations on related topics. Feel free to reach out via the email below.
What's New
April 2026
Serving as an Area Chair for NeurIPS 2026
Feburary 2026
We’re excited to announce the Embodied Reasoning in Action Workshop at CVPR 2026, where RoboSpatial will be featured as one of the official challenges - join us and submit your work!
Feburary 2026
Happy to have Watch and Learn and SpaceTools accepted to CVPR 2026!
October 2025
Watch and Learn, a framework that learns from Internet videos to enhance computer use agents, was featured in VentureBeat!
September 2025
Thrilled to see that RoboSpatial is used by Qwen3-VL and Gemini Robotics! Spatial understanding is the next frontier for MLLMs.
September 2025
Excited to organize Eval&Deploy workshop at CoRL 2025. Also attended by first robotics conference.
August 2025
Honored to serve as an Area Chair for ICLR 2026.
July 2025
Checkout Mind2Web 2, a rigorous benchmark for Deep Research and Agentic Search!
June 2025
Selected to attend CVPR Doctoral Consortium to connect with senior researchers. I'll be mentored by Prof. Katerina Fragkiadaki (CMU).
April 2025
Released Online-Mind2Web benchmark, showing current web agents are far less capable than reported.
March 2025
Interning at Google Cloud AI Research this summer working on multimodal agents. Catch me (again) in Seattle!
Feb 2025
RoboSpatial (Oral) has been accepted to CVPR 2025 with a perfect 5,5,5 score!
Feb 2025
VisualAgentBench has been accepted to ICLR 2025.
Nov 2024
Excited to present RoboSpatial, a work partly done at Nvidia. We present a large-scale 2D/3D spatial understanding dataset and benchmark tailored for robotics. Stay tuned for the full release!
Jun 2024
BioCLIP won the best student paper award at CVPR 2024! Honored to be part of the team.
Feb 2024
I will be interning at Nvidia Learning and Perception Research Group this summer. Catch me in Seattle!
Jul 2023
LLM-Planner, a paper on using large language models for vison-and-language navigation accepted to ICCV 2023.
Mar 2023
Our SalsaBot work for Amazon Alexa Prize Challenge has been accepted to the Embodied AI Workshop at CVPR 2023!
Mar 2023
I will be interning at Adobe Research this summer. Catch me in San Jose!
† indicates co-corresponding author. ‡ indicates project lead.
See full list in Publications.
-
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
Preprint Website
-
SpaceTools: Tool-Augmented Spatial Reasoning via Double Interactive RL
-
RoboSpatial: Teaching Spatial Understanding to 2D and 3D Vision-Language Models for Robotics
CVPR 2025 Oral (0.74%) / Used by Qwen3-VL , Gemini Robotics & GR00T N1.5 Paper Website Code Data
-
BioCLIP: A Vision Foundation Model for the Tree of Life
CVPR 2024 Best Student Paper Award (0.03%) Paper Website Code Data
-
LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models
ICCV 2023 Top 5 Most Cited AI Paper on arXiv (2022) Paper Website Code