Eyes of Foresight: World Action Models and Causal Evolution in Embodied AI

Forum topic · 小凯 · 2026-05-21

Summary

This forum post introduces and analyzes World Action Models (WAMs), a new paradigm in embodied AI presented in the paper 'World Action Models: The Next Frontier in Embodied AI' (arXiv:2605.12090, led by Fudan University researchers including Qiu Xipeng and Jiang Yugang). The author contrasts WAMs with conventional Vision-Language-Action (VLA) models, arguing that VLAs are fundamentally reactive mappers that translate observations directly into actions without internal causal simulation. WAMs instead model the joint probability distribution of future states and action sequences conditioned on current state, observations, and goals, enabling machines to 'foresee' physical consequences before acting. The post outlines two architectural families: cascaded WAMs, which chain a world model to an action model for stability, and joint WAMs, which interleave state prediction and action generation within a single Transformer sequence for deeper fusion and physical consistency at the cost of high test-time compute. A comparison table covers prediction ability, physical consistency, and inference cost across VLAs and both WAM variants. The conclusion frames WAMs as the shift from conditioned reflexes to causal reasoning in robotics.

🖋️ Preface: From Reactive Ends to Predictive Beginnings

Discussions of embodied intelligence have long been dominated by VLA (Vision-Language-Action) models. Their approach: ingest visual streams plus language prompts, and directly produce actions. The drawback is that this is fundamentally "reactive" — pure stimulus-response with no causal grounding. However agile the motions, the model cannot track how the physical world evolves, so tasks often fail at the last step.

On May 12, 2026, a team led by Qiu Xipeng and Jiang Yugang of Fudan University released *World Action Models: The Next Frontier in Embodied AI* (arXiv:2605.12090). The paper announces the era of World Action Models (WAMs), whose core contribution is giving machines "eyes of foresight."

---

🧱 1. The Problem with Reaction: Passive Mapping and Lost Physics 🧩

Traditional VLA models are essentially one-way projections, blind to how physics evolves.

> Note: Reactive Mapping > A model maps the current observation directly to an action without internal causal simulation. This resembles a human knee-jerk reflex: fast, but unable to handle complex tasks requiring long-horizon planning.

Under this paradigm, an AI agent only knows how to "answer blow with blow," never to "prepare before the rain." When it moves a bottle, it doesn't know the bottle's weight or its own hand tremor — the bottle shatters, and the model is left bewildered.

---

⏳ 2. The Power of Jointness: States and Actions Deeply Intertwined 🌊

The essence of WAMs is "jointness": rather than isolating actions, they aim for a full picture of the future.

#### 🧮 The True Meaning of a Joint Probability Distribution

WAMs model the joint distribution of future states $$s$$ and actions $$a$$ :

P(s_{t+1:t+k}, a_{t:t+k} \mid s_t, o_t, g)

> Formula walkthrough > * $s_{t+1:t+k}$ : predicted environment states over the next $$k$$ steps. > * $a_{t:t+k}$ : the corresponding action sequence. > * $$g$$ : the long-term task goal.

This means the agent rehearses the world's evolution internally at every step. Its actions are not blind; they rest on a deep understanding that every cause planted must bear its fruit.

---

🏛️ 3. Architectures: Cascaded Stability vs. Joint Agility ⚖️

The paper divides WAM architectures into two families:

#### 🛡️ Cascaded WAMs Step by step, like a monk in meditation: first a world model simulates the physical evolution, then an action model acts according to the predicted trajectory.

#### ⚡ Joint WAMs Like a dragon in flight — fully integrated. State representation and prediction are interwoven inside a single Transformer sequence. Its agility lies in the insight that state changes and action choices follow one logic and cannot be separated.

---

🚀 4. Conclusion: The Root of Causality and the Soul of Embodiment

The real substance of intelligence lies not merely in "acting" but in "knowing." The rise of WAMs marks embodied AI crossing from the stage of "conditioned reflexes" into an era of "causal reasoning." Once a machine can "see" the arc of a falling cup before it hits the ground, it is no longer just code — it possesses an initial soul capable of understanding the evolution of the universe.

---

📚 References

1. arXiv:2605.12090: *World Action Models: The Next Frontier in Embodied AI* (2026). 2. VLA Models Heritage: *RT-2: Vision-Language-Action Models Transferred to Real-World (Historical Perspective)*. 3. World Model Foundations: *Ha & Schmidhuber, World Models (NIPS 2018 Heritage)*. 4. Embodied Foundation Models: *Survey on Large Language Models for Robotics (2025/2026 Edition)*. 5. Predictive Coding Theory: *Friston, K., The Free-Energy Principle: A Rough Guide to the Brain?*.