The Robotics Foundation Model: From Intelligence to Action

The frontier of robot learning is shifting fast. Teams now match the performance of real-data training using largely synthetic data — and world models let machines “imagine” outcomes before acting. These advances are quietly turning robotics demos into production systems.

The session

In “From Intelligence to Action: Building the Foundations of Physical AI,” Jinwoo Shin lays out the building blocks of a robotics foundation model — how robots learn to reason about the physical world and turn that reasoning into action.

Why it matters now

Recent results show robots trained on 40% synthetic data matching policies trained entirely on real data. For an industry bottlenecked by the cost of real-world data collection, that's a turning point — and it's reshaping how the next generation of robots will be built.

About the speaker

Jinwoo Shin is a Professor at KAIST and one of Korea's leading machine-learning researchers, working at the core of robotics foundation models, world models, and the learning methods behind physical AI.

Key Takeaways

1. World models — learned simulators that let robots plan and evaluate actions before executing them.
2. Vision-language-action — connecting perception, language, and control into a single loop.
3. Efficiency over scale — why well-curated data now beats raw size in robotics foundation models.

Session at a Glance