Generative AI Talent Development Program

Workforce program for generative AI; KETI's research develops Vision-Language-Action (VLA) models for robot manipulation

KETI participates in the Generative AI Talent Development program (led by Sejong University), which pairs next-generation AI workforce development with core research. KETI’s research thread develops Vision-Language-Action (VLA) models for human-motion imitation — generating general-purpose robot tasks with stepwise reasoning.

VLA model for human-motion imitation: an Inverse Dynamics Model supplies latent-token action representations and a VLM handles low-level task planning, feeding a diffusion transformer (skill-space embedding) and action decoder for general robot task generation.

Research approach

  • Latent-skill-based hierarchical imitation learning with few-shot adaptation
  • An Inverse Dynamics Model (IDM) extracts latent-token action representations and a Vision-Language Model (VLM) performs low-level task planning, both feeding a diffusion transformer (skill-space embedding) that drives an action decoder
  • We are currently investigating uncertainty quantification (UQ) to enhance the VLA dataset — using UQ to improve the quality and selection of demonstration data

Funding: IITP (Ministry of Science and ICT)
Period: 2025 –
Role: Participant (KETI)