How Generative World Models Like Sora Are Scaling Robot Physical Intelligence and Accelerating the Path to AGI - 기기픽

🤖 AI In-Depth Analysis Report

This post is a future-oriented analysis report generated by an AI model (Gemini), based on global AI research trends and large-scale data from the United States, Europe, and beyond.

This is not a simple translation of a specific paper. Instead, it presents original insights synthesized and evaluated autonomously by AI. Experience a fresh perspective rarely found in existing domestic discussions.

---

How Generative World Models Like Sora Are Scaling Robot Physical Intelligence and Accelerating the Path to AGI

When OpenAI’s Sora was first unveiled, the public was captivated by its stunning video quality. Robotics researchers, however, felt a very different kind of shock—one rooted in the realization that a long-standing bottleneck in robotics might finally have a solution: the data scarcity problem in physical intelligence.

Until now, robotics has been constrained by Moravec’s Paradox. While AI can outperform humans at chess, it struggles with seemingly trivial tasks such as walking or grasping a cup. The fundamental reason lies in data scalability. Text data is effectively infinite on the internet, but data from robots interacting with the real world is extremely limited.

In this column, we take a deep dive into how generative world models (such as Sora) are unlocking scalable data strategies for robot physical intelligence, and why Video-to-Action technology is emerging as a core tech trend for 2025.

---

1. The Greatest Bottleneck in Robotics: Data Famine

The success formula for large language models (LLMs) was straightforward: “Add more data and more compute power”—the scaling law. However, this formula has not translated well to Embodied AI.

• Limits of real-world data collection: Training robots by physically breaking cups and repeating tasks is expensive, slow, and risky.
• The Sim-to-Real gap: Traditional physics-engine simulators fail to capture real-world complexity—friction, lighting changes, material deformation, and more.

For robots to reach human-level physical intelligence, they require massive-scale data that internalizes real-world physics. This is where generative AI—specifically, world models—enters the picture.

2. Generative World Models: Digital Imagination for Robots

Viewing video generation models like Sora as mere content creation tools misses the point. At their core, they function as general-purpose physics simulators.

Three Key Advantages World Models Offer Robots

1. Infinite training environments: Robots can run millions of trials in generated video worlds, learning causal relationships— such as “if a cup falls, it breaks.”
2. Future prediction capability: By generating the next frame, a world model can answer: “What will happen if I move my arm this way?”
3. Internalization of physical laws: Without explicit equations, models learn gravity, inertia, and collision dynamics directly from large-scale video data.

3. Core Strategy: Video-to-Action Pre-training

How, then, do we translate video into robotic action? This question lies at the technical heart of scaling robot physical intelligence with generative world models.

Breaking Down the Video-to-Action Pipeline

This strategy consists of two main stages, mirroring the pre-training → fine-tuning paradigm used in LLMs.

• Step 1: Internet-scale video pre-training: Billions of online videos teach the model how the world moves. The world model learns physical context by predicting subsequent frames (e.g., V-JEPA, Sora).
• Step 2: Action labeling via inverse dynamics models: The model infers what action was required to transition from one state to the next— learning (current state, next state) → action.

Through this approach, robots can acquire general physical intelligence— learning, for example, how to move a wrist for cutting— without ever physically manipulating a robot arm.

4. Key Players and Market Outlook (Project GR00T, Figure AI)

Major technology leaders are already engaged in this data scalability race.

• NVIDIA (Project GR00T): Building foundation models for humanoid robots by combining generative AI with simulation platforms like Isaac Lab.
• Google DeepMind (Genie): Creating interactive virtual worlds from game videos—enabling infinite environments for robot learning.
• Figure AI & OpenAI: The Figure 01 robot, powered by OpenAI’s multimodal models, demonstrates advanced perception-to-action capabilities.

5. Conclusion: The “ChatGPT Moment” for Robotics Is Approaching

Scaling robot physical intelligence through generative world models is more than a trend—it represents a paradigm shift that dismantles decades-old data limitations in robotics.

Robots will no longer rely solely on scarce lab data. Instead, they will learn from humanity’s collective video corpus. Just as GPT transformed language, foundation models for the physical world will redefine robotic generality.

Key Takeaways:

• World models are physics simulators, not just video generators.
• Video-to-Action is the most viable solution to robotic data scarcity.
• Post-2025, control over physical intelligence will define tech leadership.