Frontier Insights
Technical explainers written by the project team and collaborating researchers — pick what fits your role.
Dyna: A Model That Teaches Robots to Act via "Video Prediction"
Dyna represents a World-Action Model approach that leverages video diffusion models to simultaneously predict future frame evolutions and robot actions from vast human egocentric videos, learning physical world dynamics to overcome traditional teleoperation data bottlenecks and achieve highly robust, cross-platform robot manipulation with minimal fine-tuning.
Wan-Chi Chang · Department of Computer Science, National Tsing Hua University
Agent Skills: Teaching AI Reusable Professional Capabilities
Agent Skills encapsulate workflows, rules, and domain context into reusable AI capabilities, enabling models to perform complex tasks through progressive disclosure without cluttering context windows.
Wei-Chen Huang · Department of Statistics, National Taipei University (NTPU)
VLA: Teaching Robots to Understand Plain Language
Bringing the recipe behind language models into the physical world — how VLA lets robots handle objects and instructions they have never seen before.
Da-An Li · Department of Computer Science, National Tsing Hua University