Suhwan Choi

I'm an undergraduate student majoring in Physics and Computer Science at Seoul National University. I'm currently a Principal Researcher at Maum.ai, where I lead the autonomous robotics research division.

My main research interests are in approximating and imitating human behavior and intelligence in multimodal modalities, utilizing end-to-end architectures and scalable training suites. I focus on embodied AI, robotic navigation, vision-language models, and multimodal learning.

Email  /  CV  /  LinkedIn  /  Github  /  Blog

Research & Publications

I work on embodied AI, robotic navigation, and multimodal learning. My research focuses on scaling vision-action pretraining, commonsense-aware navigation systems, and vision-language model improvements. Some papers are highlighted.

An interactive gaming world model using camera pose as a unifying geometric representation for precise action control and long-horizon 3D consistency.

A navigation benchmark that shifts evaluation from purely technical metrics to real-world economic cost and revenue, featuring high-fidelity physics simulation with delivery robot dynamics.

Revisiting residual connections with orthogonal updates for more stable and efficient deep networks.

Exploiting semantic reconstruction to mitigate hallucinations in vision-language models.

Experience

• Founded autonomous robotics research division as the first researcher, leading strategic decisions and team expansion to 10 researchers.
• Contributed as first author to majority of research projects in robotic navigation and embodied AI.
• Led CORE: Slurm-based DGX Cluster construction project (96 H100 GPUs, 12 nodes). [Blog]
• Implemented company-wide Notion workspace enhancing productivity and streamlining workflows. [Template]

• Worked on diffusion-based personalized profile image generation for real-world applications.

Open Source Projects

• Primary author of a unified evaluation framework for Vision-Language-Action models across robot simulation benchmarks. Features batch parallel evaluation with 47x throughput, Docker-isolated benchmarks, and the largest unified VLA leaderboard (500+ models x 17 benchmarks).

• A structured reading list on Vision-Language-Action (VLA) models — from diffusion/flow matching foundations through state-of-the-art robot foundation model architectures to data scaling, RL fine-tuning, and world models.

• Unofficial reimplementation of VLA-0 using TRL's SFTTrainer. While common VLA codebases are over 10,000 lines, vla0-trl contains only ~1,200 lines total. Gets ~92% on LIBERO by just fine-tuning Qwen2.5-VL to predict actions as text, without any custom architecture.

• WebSocket-based RPC library for Python. Uses transparent proxy objects to interact with remote Python objects as if they were local, with convenient module patching for minimal code changes.

• A lightweight Python library for lazy-loading registries with namespace support and type safety.

• Pydantic media reference for images and video frames (with timestamp support) from data URIs, HTTP URLs, file URIs, and local paths. Features lazy loading and optimized batch video decoding.

• Real-time Slurm cluster monitoring tool with interactive TUI with Textual. Visualizes GPU/CPU/memory allocation across nodes with job-level drill-down.

Core contributor and maintainer with 180+ merged PRs. Built comprehensive multimodal desktop agent framework including optimized data collection tool (ocap), standardized efficient data format (OWAMcap), dataset visualizer, multimedia data management/processing pipelines, agent training, Python packaging, and CI/CD infrastructure.

Awards & Honors

• Ranked 4th/793 teams in 2023, Ranked 3rd/618 teams in 2024.
• Contest implementing quantum algorithms, quantum machine learning, quantum chemistry, and brain-teasing puzzles.

• 1st place, Minister of Science and ICT Award
• Topic: Utilizing symmetry to solve variational quantum algorithm (quantum machine learning) efficiently.

• 3rd place on Landmark Detection (3,000,000 KRW)
• 2nd place on Shopping User Embedding Extraction, Classification (7,000,000 KRW)

• Round 3, 546th (awarded T-Shirt).

Website source code available on GitHub.