LV Robotics: Foundation Models Are Changing How We Build Robots

About This Event

For most of robotics history, intelligent automation was built forward: Task → Program → Robot → Action. Whenever a robot needed to perform a new task, engineers had to write new code, gather fresh demonstrations, retrain the architecture, and perform painstaking real-world tests. Today, that paradigm is shifting dramatically. A new generation of robot foundation models, Vision-Language-Action (VLA) models, and world models is decoupling physical intelligence from specific tasks, and even from specific hardware embodiments.

Breakthroughs We’ll Explore

• Google DeepMind’s Gemini Robotics 2: Leveraging unified models across diverse robot embodiments and rapidly adapting to unencountered physical platforms.
• Stanford’s SimToolReal: Achieving zero-shot transfer of complex, dexterous tool manipulation directly from simulation to real-world environments.
• Physical Intelligence’s π0.7: Demonstrating robust zero-shot cross-embodiment generalization and immediate performance on novel tasks.
• OS3: Pretraining physical intelligence across 150,000 hours of video content spanning over 2,000 distinct tasks.
• Moving Atoms: Building action-conditioned world models that enable robots to generate and learn from synthetic physical experience on GPUs.
• Shiraz AI: Developing foundation models that translate human video demonstrations into actionable robot behaviors across embodiment gaps.

From Forward to Reverse Inference

We are moving away from traditional forward inference (Job → Training → Model → Action) toward reverse inference: evaluating a model's foundational physical capabilities to instantly infer what jobs a given robot can perform. This transition unlocks powerful possibilities:

• Simulation becomes a primary source of experience.
• Human video acts as scalable training data.
• Skills become highly transferable across interchangeable hardware bodies.

Key Discussion Topics

• What defines a true robot foundation model vs. standard VLA?
• How close are we to seamless, zero-shot sim-to-real transfer?
• Can a single universal model effectively control disparate robotic hardware?
• Are we reaching the point where robots infer tasks simply by observing the world?

Join the Conversation!

Whether you are a roboticist, AI researcher, hardware engineer, or tech enthusiast, join us for an engaging discussion on one of the most critical evolutions in modern robotics! RSVP now to reserve your spot and connect with the LV Robotics community.