Physics First: Why Hamiltonian Mechanics Is the First Principle of Digital-Twin Brains

Forum topic · 小凯 · 2026-05-21

Summary

A Chinese tech forum post argues that pure data-driven world models suffer from 'physical drift'—objects vanish, gravity inverts, and energy explodes when video-generation models are deployed in robotic control. It highlights a paper titled 'Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling' (arXiv:2605.00412), which proposes the Hamiltonian World Model (HWM). HWM upgrades the latent space to a 'latent phase space' with paired coordinates: generalized position q and momentum p, constrained by symplectic geometry so predictions stay stable over long horizons. The model adds three corrections to ideal energy-conserving physics: a control term injecting robot actions into phase space, a dissipation term capturing friction and drag, and a neural residual term for hard-to-formalize forces like collision deformations. Reported advantages over VAE/Transformer world models include converging with thousands instead of hundreds of thousands of interactions, stable rollouts beyond 100 steps versus collapse after 10, and interpretable latent variables with physical meaning. The author frames physics-native modeling as a foundation for AGI and robotics.

Preface: When 'Sora-Style Hallucinations' Meet Real Gravity

Between 2024 and 2025, the world was stunned by a wave of impressive video generation models. We watched coffee swirl in cups and pedestrians stroll Tokyo streets, as if a true "world simulator" were within reach.

Yet when these models were sent into labs and installed as the "brains" of robotic arms, a massive reveal took place. The world inside a robot's mind is often bizarre: objects vanish into thin air, gravity suddenly reverses, and energy explodes for no reason. This "Physical Drift" is the fatal flaw of all purely data-driven models—they understand pixels, but not laws.

Until, on May 1, 2026, Sen Cui and Jingheng Ma published a major paper on arXiv: "Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling" (arXiv:2605.00412), pointing out the real way forward. They argue that to build a true digital-twin brain, we must return to the starting point of physics: Hamiltonian Mechanics.

---

1. The Hamiltonian World Model (HWM): Carving Newton's Laws into Latent Space

Traditional AI world models grope in the dark, trying to "intuit" physics from massive video data. HWM instead hands AI a physics instruction manual.

> Note: The Hamiltonian ( $$H$$ ) > In classical physics, the Hamiltonian represents the total energy of a system. If AI can learn a Hamiltonian function consistent with energy conservation, its predicted world will never produce absurd scenarios like perpetual motion machines or objects vanishing instantly.

HWM upgrades the traditional AI "latent space" into a physics-grade "Latent Phase Space." Here, every state is no longer an isolated point but a precise pairing of twin variables:

$$q$$ (generalized coordinates): Where is the object?
$$p$$ (generalized momentum): Where is it headed?

This paired structure ensures the system's evolution obeys the topological constraints of Symplectic Geometry. Simply put, it keeps the AI's predicted trajectories stable like planetary orbits—still on track even after ten thousand rollout steps.

---

2. Control, Dissipation, and Residuals: Bringing Ideal Physics Down to Earth

The real world is imperfect. With only ideal energy conservation, a robot cannot simulate friction or motor output. HWM's brilliance lies in three key correction terms:

#### Control It explicitly injects the robot's actions into phase space via a control matrix. This lets AI not only "see" the world but also predict "if I push, where will the ball roll."

#### Dissipation It models real-world energy loss:

\frac{dp}{dt} = -\frac{\partial H}{\partial q} - D(q, p)

> Note: Dissipation Term ( $$D$$ ) > It captures forces like friction and air resistance that slowly bleed away energy. Without it, the AI's simulated pendulum would swing forever.

#### Residual This is the "compromise" between AI and physics. For extremely complex forces that are hard to formalize (e.g., microscopic deformation during collisions), the model uses a neural network's generalization power to fill the gap.

---

3. A Leap in Data Efficiency: No More Blind Learning

In robotic experiments, HWM delivers a crushing advantage:

Because HWM has physics built in, it doesn't need to learn from scratch that "balls fall" due to gravity. It only needs a small amount of data to align parameters, then can precisely simulate complex dynamics in its "mind."

---

4. Conclusion: The Physical Foundation on the Road to AGI

I firmly believe: true intelligence must be anchored in the real physical world.

Where BrainDyn reveals the continuity of brain dynamics, HWM provides the solid skeleton needed to build "silicon-based life." When we stop piecing the world together from pixels and instead understand all things through energy and momentum, we cross the deepest chasm on the path to AGI—the one named "reality."

---

References

1. arXiv:2605.00412: *Physically Native World Models: A Hamiltonian Perspective on Generative World Modeling* (2026). 2. Hamiltonian Neural Networks: *Greydanus et al., Hamiltonian Neural Networks (NeurIPS 2019 Heritage)*. 3. Symplectic ODE Solvers: *Symplectic Integration for Generative Modeling in Robotics (2025 Archive)*. 4. Generative World Models: *Ha & Schmidhuber, World Models (2018/2026 Evolution)*. 5. Robot Learning Frameworks: *Data Efficiency in Model-Based Reinforcement Learning*.