Research Intern, World Model

Project Overview

WoRV is developing a World Foundation Model (WFM) for indoor multi-environment mobile manipulation robots as part of the "Physical AI Leading Technology Development" national R&D consortium. We train our own VLA (Vision-Language-Action) model that integrates video, language, and action — and we close the entire cycle from top-tier conference publication to real-robot demonstration and share it publicly.

To support this, we have on-premise designed and built our own 96 H100 GPU high-performance training cluster and a 600+ TB · 200 GB/s parallel storage system, and we are building a multimodal training dataset of 10,000+ hours through our own collection pipeline based on UMI rigs and bimanual robot platforms. On top of this infrastructure, we are building world models at the frontier of the field.

We are looking for passionate researchers to join this project. We welcome applications from anyone interested in World Model / VLA research and development.

WoRV is a fast-growing in-house startup. The process bends to fit the role and the candidate — we'd rather get to know you well than push you through a fixed funnel.

Research Topics

The World Model research team works at the intersection of the following topics. Interns participate deeply in one track based on their interests and skills; the match is determined together during the technical interview.

• Long-context multimodal modeling — learning and inference architectures for long-horizon video, language, and action sequences
• Policy learning for real-robot deployment — models that simultaneously satisfy latency, robustness, and generalization requirements
• VLA / Robotics Foundation Model training pipelines — pre-training, fine-tuning, and evaluation cycle operations
• Closed-loop and OOD evaluation — quantifying generalization performance on simulation and real-robot benchmarks
• Scaling analysis & data curation — training data mixture design and scaling law validation

Focus topics may shift as the field evolves and research progresses — we concentrate resources on the highest-impact direction at each moment.

Responsibilities

• VLA model training and large-scale experiment management
• Experiment design, ablation studies, and result analysis
• Reproducing public benchmarks and using evaluation frameworks such as vla-eval (co-developed with AI2)
• Closed-loop evaluation in simulation and on real robots
• Co-authoring top-tier conference papers (co-authorship possible based on contribution)

Requirements

• Practical deep learning research and implementation experience (PyTorch, project-level or above)
• Multi-GPU distributed training experience
• Ability to design ablations and manage experiment reproducibility
• Ability to read and write English academic papers
• Available for full-time 3+ months

Preferred Qualifications

• Experience training or reproducing VLA / VLM models (π0, GR00T, etc.)
• Research experience in robot policy learning: Imitation Learning, Reinforcement Learning, etc.
• Experience implementing Diffusion / Flow Matching policy models
• Experience designing training data mixtures or validating scaling laws
• First-author publication at a top-tier venue (NeurIPS, ICLR, CVPR, ICRA, IROS)

Work Arrangement

• Full-time 3 months (extension or conversion to full-time negotiable)
• Second Pangyo IT Center (on-site)
• Competitive internship compensation
• Priority consideration for full-time Research Engineer position after the internship

Research Infrastructure

• On-premise 96 H100 GPU training cluster (CORE Cluster) — in-house designed and built, research-dedicated, no queue
• In-house distributed storage cluster — 600+ TB · 200 GB/s bandwidth
• 10,000+ hours of multimodal training data
• In-house data collection pipeline (UMI rigs, bimanual robot platforms, etc.)
• Weekly 1:1 mentoring, paper reading, and research sync
• Full sponsorship for international conference presentations

Application Notes

• Resume + cover letter (PDF)
• Portfolio (GitHub, papers, project links)
• Select 1–2 research topics of interest and write 1–2 paragraphs on why (in the Google Form's Additional Self-PR section)