My profile image

Yongjun Cho

My research sits at the intersection of learning and embodiment — I am drawn to the question of how machines can acquire not just knowledge, but the skills to act reliably in the world.

Over 4+ years, I have built and deployed autonomous navigation systems end-to-end — developing VLA pipelines (1B–7B params) deployed to 4 field sites at a 97.3% success rate, and shipping models to edge hardware (Jetson Thor, FP8). The work has been recognized at IROS 2022, ICRA 2025, and ICLR 2026, including a NeurIPS 2024 Workshop Outstanding Paper (top 3 of 100), and has led to 2 enterprise contracts and my selection as lead developer for a $25M government foundation model project.

Research Interest: Robotics, Autonomous Navigation, Vision-Language-Action Models, Sim-to-Real Transfer, Training Pipelines, On-Device Inference, Reinforcement Learning, Diffusion Models

✉️ cyjun0304@kaist.ac.kr

CVLinkedInGoogleScholarGitHub

News

2026-01-26

🎉 Our paper D2E got accepted to ICLR 2026! Bridging desktop data to embodied AI - scaling vision-action pretraining with 1.3K+ hours of data achieving 96.6% success on LIBERO! 🤖

More

2025-01-28

🎉 Our paper CANVAS got accepted to ICRA 2025!

More

2024-12-15

🏆 Outstanding Paper Award at NeurIPS 2024 Open-World Agents Workshop (3 selected out of 100 accepted papers).

More

Publications

† denotes equal contribution, * denotes corresponding author

CostNav: A Navigation Benchmark for Cost-Aware Evaluation of Embodied Agents

Co-First Author

[0]Haebin Seong†, Sungmin Kim†, Yongjun Cho†, , Youngjae Yu*, Yunsung Lee*

CostNav: A Navigation Benchmark for Cost-Aware Evaluation of Embodied Agents

arXiv preprint 2026

[arxiv][blog][code]

EditCrafter: Tuning-free High-Resolution Image Editing via Pretrained Diffusion Model

[1]Kunho Kim, Sumin Seo, Yongjun Cho,

EditCrafter: Tuning-free High-Resolution Image Editing via Pretrained Diffusion Model

CVPR 2026 2nd Workshop on Human-Interactive Generation and Editing (HiGen) Proceeding Track

[website][arxiv][code]

D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI

[2]Suhwan Choi†, Jaeyoon Jung†, Haebin Seong†, , Youngjae Yu*, Yunsung Lee*

D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI

The Fourteenth International Conference on Learning Representations (ICLR) 2026

[website][arxiv][code]

CANVAS: Commonsense-Aware Navigation System for Intuitive Human-Robot Interaction

Co-First Author

[3]Suhwan Choi†, Yongjun Cho†, Minchan Kim†, , Youngjae Yu*

CANVAS: Commonsense-Aware Navigation System for Intuitive Human-Robot Interaction

International Conference on Robotics and Automation (ICRA) 2025

NeurIPS 2024 Workshop Open-World Agents (Outstanding Paper Award)

[website][award][IEEE]

Fine-tuning Pocket-conditioned 3D Molecule Generation via Reinforcement Learning

Co-First Author

[4]Daeseok Lee†, Yongjun Cho†

Fine-tuning Pocket-conditioned 3D Molecule Generation via Reinforcement Learning

ICLR 2024 Workshop on Generative and Experimental Perspectives for Biomolecular Design (GEM)

[openreview][code]

Sim-to-Real Transfer of Image-Based Autonomous Guidewire Navigation Trained by Deep Deterministic Policy Gradient with Behavior Cloning for Fast Learning

Co-First Author

[5]Yongjun Cho†, Jae-Hyeon Park†, Jaesoon Choi, , Dong Eui Chang*

Sim-to-Real Transfer of Image-Based Autonomous Guidewire Navigation Trained by Deep Deterministic Policy Gradient with Behavior Cloning for Fast Learning

2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

[IEEE]

Image Processing Based Autonomous Guidewire Navigation in Percutaneous Coronary Intervention

First Author

[6]Yongjun Cho, Jae-Hyeon Park, Jaesoon Choi, , Dong Eui Chang*

Image Processing Based Autonomous Guidewire Navigation in Percutaneous Coronary Intervention

2021 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia)

[IEEE]

Real-time Quadrotor Actuator Fault Detection and Isolation Using Multivariate Statistical Analysis Techniques with Sensor Measurements

[7]Jae-Hyeon Park, Yongjun Cho, Jin-Yeong Jeong, , Dong Eui Chang*

Real-time Quadrotor Actuator Fault Detection and Isolation Using Multivariate Statistical Analysis Techniques with Sensor Measurements

2020 20th International Conference on Control, Automation and Systems (ICCAS)

[IEEE]

Experience

Senior Research Scientist @ Maum AI

Seongnam, South Korea · Nov. 2024 - Present

  • Deployed VLA navigation models (1B params) to 4 real-world sites, improving success rate from 75% to 97.3% through human intervention data collection and sim-to-real domain randomization.
  • Built multi-node training infrastructure: 5x storage reduction (video compression), 2x I/O throughput (Ceph), 3x training speedup (DeepSpeed ZeRO Stage 2, 4+ nodes, Slurm).
  • Quantized VLA model to FP8 and deployed on NVIDIA Jetson Thor, collaborating with on-device deployment team on ROS2 inference pipeline.
  • Ran technical PoCs and field demos at orchards and golf courses, directly leading to 2 enterprise contracts with major Korean corporations.
  • Built benchmark environments and evaluation pipelines for D2E — 1,300+ hour pretraining data pipeline, open-sourced with 8,200+ downloads.
  • Designed and open-sourced a navigation benchmark with economic cost modeling (simulation + 50-hour dataset + eval tools). Supervised SNU research intern.
  • Selected as lead developer of the World Foundation Model for a government national project (IITP, 340B KRW / $25M budget, 2026–2027).

Research Scientist @ Maum AI

Seongnam, South Korea · May 2024 - Oct. 2024

  • Designed CANVAS, a VLA navigation framework controlled via sketches and natural language — published at ICRA 2025, awarded NeurIPS Workshop 2024 Outstanding Paper (top 3 of 100).
  • Built the COMMAND dataset (48 hours, 219 km of real-world driving) and multi-container inference architecture used in all subsequent VLA deployments.
  • Designed evaluation protocols in NVIDIA Isaac Sim and real-world environments, establishing the data collection and sim-to-real pipeline for the team.
  • Filed 2 Korean patents on robot control via sketch and natural language commands.

Machine Learning Researcher @ Deargen Inc.

Seoul, South Korea · Apr. 2022 - Apr. 2024

  • Built end-to-end Drug-Target Interaction pipeline (crawling, model design, training, eval) and delivered a custom fine-tuning service to a global big pharma client.
  • Designed PPO reward shaping for pocket-conditioned 3D molecule generation — published at ICLR 2024 Workshop, filed as Korean patent.

Education

M.S. in School of Electrical Engineering

KAIST (Korea Advanced Institute of Science and Technology) · Daejeon, S.Korea · Mar. 2020 - Mar. 2022

  • Advisor: Dong Eui Chang
  • Built an image-based autonomous guidewire navigation system (DDPG + Behavior Cloning) on a custom-built vascular phantom testbed, achieving Sim-to-Real transfer (IROS 2022).
  • Won the Grand Prize at the DAPA Autonomous Drone Competition (20+ teams) by developing 2D autonomous exploration; deployed on Jetson TX2 via Sim-to-Real transfer from Gazebo.
  • Led HD Korea Shipbuilding industry collaboration on ship-assistance drones, building GPS-based navigation with relative path planning.
  • Developed a real-time multivariate statistical anomaly detection system for quadrotor actuators (ICCAS 2020).

B.S. in Department of Mechanical Engineering

KAIST (Korea Advanced Institute of Science and Technology) · Daejeon, S.Korea · Mar. 2015 - Mar. 2020

  • Graduation project: built an autonomous driving robot for library inventory management.

Honors & Awards

Outstanding Paper Awards

NeurIPS 2024 Workshop Open-World Agents

Vancouver, Canada

International

2024

Grand Prize

Autonomous Drone Competition hosted by the Defense Acquisition Program Administration

Daejeon, S.Korea

Domestic

2020

National Excellence Scholarship (Natural Sciences and Engineering)

Covers admission fee, full tuition, and additional support for study grant and living expenses

Korea Student Aid Foundation (KOSAF)

Scholarship

2015 - 2018