VLA Is Not Dead—At Least Not Yet

说 VLA 已死，为时过早

这篇 DreamZero / WAM 论文最重要的点，不是“机器人会生成视频了”，而是它把 robot policy 重新拆成了两个问题：先预测未来会发生什么，再从这个未来反推出动作。

论文 Eq. 1 已经把结构写得很清楚：

先生成 future video，再在这个 future video 上生成 action。也就是说，WAM 的本质是：显式 forward dynamics model + IDM。

它不是传统 VLA 那种从 VLM 接一个 action head，直接做 image / language → action。它先生成 future video；再让 inverse dynamics model 根据这个 future state 生成动作。future video 在这里扮演的是 visual plan。IDM 做的事情，是把这个 visual plan 翻译成 robot action。

这个区别非常关键。

为什么 WAM 在数据上更占便宜

VLA 的预训练主要来自 image-text / VLM 数据。它学到的是语义常识：

• 这个东西是什么？
• 这句话是什么意思？
• 这个目标大概在哪里？

但机器人真正缺的不是这个，而是：这个动作做下去，物理世界会怎么变？

WAM 的 advantage 在这里。

video generation backbone 本身就预训练在大规模视频上。视频天然包含时间、运动、接触、遮挡、变形、因果演化。虽然互联网视频没有 robot action label，但它至少和机器人最终要预测的东西高度对齐：都是物理世界的状态序列。

所以 robot data finetune 的阶段，本质不是从零学物理，而是让一个已经见过大量世界演化的视频模型，学会：

• 这个 embodiment 怎么和世界互动；
• 这个 future visual trajectory 对应什么 motor command。

相比之下，VLA 的 VLM pretraining 和 robot action prediction 是错位的。image-text data 对“语义”很强，但对“动作序列的物理演化”并不直接对齐。

这也是为什么 WAM 能更好吃下 diverse、non-repetitive robot data。对于 VLA，同样一段杂乱 robot trajectory 可能只是 noisy state-action pairs。对于 WAM，每一段连续视频都是 dense future-state supervision。

这不是小差别，这是数据利用方式的差别。

为什么 IDM 比直接学 VLA policy 更好学

直接 VLA policy 要一次性学两个东西：

• 未来应该发生什么；
• 机器人该怎么动才能让它发生。

WAM 把问题拆开：

• video model 负责预测 future state；
• IDM 负责把 future state 映射成 action。

这会让 action learning 更像一个逆动力学问题，而不是端到端硬拟合 policy。

论文里也给了很强的信号：很多失败主要来自 video prediction error，而不是 action extraction error。换句话说，只要未来视频预测对了，IDM 往往能比较稳定地把动作拉出来。

这说明 WAM 的瓶颈很可能已经从“动作头会不会学”转移到了“世界模型能不能预测对”。这也是为什么它在 unseen task 和 cross-embodiment 上效果好，是合理的。

为什么 unseen task / cross-embodiment 结果 make sense

对于 unseen task，VLA 如果没在 robot data 里见过类似动作，它只能靠 VLM 的语义 prior 硬猜。但 VLM 没有大规模动作序列预训练，它并不知道“解鞋带”“熨衣服”“握手”这种物理 motion 应该如何展开。

WAM 不一样。

它可以先用 video generation backbone 生成一个 future visual plan。这个 video backbone 很可能在大规模视频里见过类似的动作演进。然后 IDM 再把这个 visual future 转成 robot action。

所以 WAM 在 unseen task 上比 VLA 好，不神秘。

cross-embodiment 也是一样。如果一个 human 或另一个 robot 的 video-only demo 能帮助 WAM 更新“这个任务的视觉演化”，那它不一定需要 action label。因为 action label 是 embodiment-specific 的，但 future visual state 更接近 embodiment-agnostic。

这就是 WAM 最有价值的地方：它把大量没有 robot action label 的视频数据，变成了可以影响 policy 的训练信号。

这是 VLA 很难直接做到的。

但代价也很大

WAM 的强，来自显式 world modeling。但显式生成 future video 不是免费的。

它需要模型真的理解多出来的世界演化信息。这就吃 model capacity。

Table 4 里这个点非常明显：

5B DreamZero 只有 21% task progress；14B DreamZero 到 50%。

小模型扛不住显式 video generation + action alignment，容易 hallucinate future，然后 action 也跟着错。

所以 WAM 是一个把 world model 放进 policy 里的大模型系统。这和很多 VLA 的工程属性不一样。VLA 可以更便宜、更快、更容易部署；WAM 则把更多能力前置到了 video generation backbone 里。

真正尴尬的是工业落地

如果只看 zero-shot / low-shot，WAM 很漂亮。但如果你真要落地到工厂、仓库、餐饮、家庭服务，一个绕不开的问题是：

最终还是要 post-train。

Figure 10 很值得细看。

在 shirt folding 上，33 小时 post-training data 之后，DreamZero 和 pretrained VLA 基本打平，都是 92.5%。在 fruit packing 上，12 小时 post-training data 之后，DreamZero 到 96%，明显强于 VLA。

这就带来一个很尴尬的 tradeoff：

少量数据时，WAM 确实比 VLA 更强，但可能还没强到足够稳定落地；数据加多之后，WAM 和强 VLA 的差距又可能缩小；但 WAM 的推理成本、延迟、系统复杂度明显更高。

论文自己也承认，DreamZero 要靠一整套系统优化，把 14B autoregressive video diffusion model 从 5.7 秒优化到 150 ms，并且使用 2×GB200 才跑到 7 Hz。一台 GB200 的成本在 50 万元以上，而且需要直接液冷（DLC）系统。与此同时，当前 VLA 可以在 consumer GPU 上跑到 20 Hz+。

这不是小工程差距。这是部署成本差距。

所以，如果一个工业任务本来就可以收几十小时高质量数据，VLA 便宜、快、稳定，可能依然是非常强的 baseline。

WAM 的优势更像是：

• 长尾任务；
• unseen motion 及低重复示教下的泛化；
• cross-embodiment transfer；
• video-only data 利用。

而不是所有场景直接替代 VLA。

我的判断

WAM 是一条非常重要的路线。它的核心贡献不是“机器人生成视频”，而是证明了：

video generation backbone 可以作为 robot policy 的物理先验。

这条路线比传统 VLA 更接近物理世界的演化结构，也更能利用大规模视频数据。它让 robot learning 从“直接拟合动作”变成了“先理解未来，再反推动作”。

这很强。但说 VLA 已死，为时过早。

VLA 的问题是物理泛化弱；WAM 的问题是成本、延迟、容量和部署复杂度高。

一个更现实的判断是：

WAM 会在 unseen task、low-shot、cross-embodiment 和 video-data scaling 上打开新上限；但在大量工业场景里，VLA 仍然会因为便宜、快、易部署而活得很好。

WAM 证明了 VLA 的上限问题。但它还没有证明 VLA 的工程价值已经消失。

所以，VLA 没死。只是以后不能再只靠 VLM pretraining 讲 robot foundation model 的故事了。

VLA Is Not Dead—At Least Not Yet

The most important point in the DreamZero / WAM paper is not that “robots can generate video now.” It is that the paper decomposes a robot policy into two problems: first predict what will happen in the future, then infer the action from that future.

Equation 1 makes the structure explicit:

It first generates a future video and then generates actions on top of that video. In other words, the essence of WAM is an explicit forward dynamics model + IDM.

This is different from a conventional VLA that attaches an action head to a VLM and directly maps image / language → action. WAM first generates a future video; an inverse dynamics model then produces actions from that future state. The future video serves as a visual plan, and the IDM translates that plan into robot actions.

That distinction is crucial.

Why WAM Has a Data Advantage

VLA pretraining mainly comes from image-text / VLM data. It learns semantic common sense:

• What is this object?
• What does this sentence mean?
• Where is the target likely to be?

But what robots truly lack is something else: if I take this action, how will the physical world change?

That is where WAM has an advantage.

A video-generation backbone is already pretrained on large-scale video. Video naturally contains time, motion, contact, occlusion, deformation, and causal evolution. Internet video may not have robot-action labels, but it is closely aligned with what a robot ultimately needs to predict: a sequence of states of the physical world.

Robot-data fine-tuning therefore does not have to teach physics from scratch. Instead, it teaches a video model that has already seen extensive world dynamics:

• how this embodiment interacts with the world; and
• which motor commands correspond to a future visual trajectory.

By contrast, VLM pretraining and robot-action prediction are misaligned in a VLA. Image-text data is strong on semantics, but it is not directly aligned with the physical evolution of action sequences.

This also explains why WAM can make better use of diverse, non-repetitive robot data. For a VLA, the same messy robot trajectory may look like noisy state-action pairs. For WAM, every continuous video segment provides dense future-state supervision.

This is not a minor difference. It is a fundamentally different way of using data.

Why an IDM Is Easier to Learn Than a Direct VLA Policy

A direct VLA policy must learn two things at once:

• what should happen in the future; and
• how the robot should move to make it happen.

WAM separates them:

• the video model predicts the future state; and
• the IDM maps that future state to an action.

This turns action learning into something closer to an inverse-dynamics problem, rather than forcing an end-to-end policy fit.

The paper also gives a strong signal: many failures come primarily from video-prediction errors rather than action-extraction errors. In other words, once the future video is correct, the IDM can often recover the action reliably.

WAM may therefore have moved the bottleneck from “can the action head learn?” to “can the world model predict correctly?” That makes its strong performance on unseen tasks and cross-embodiment transfer quite plausible.

Why the Unseen-Task and Cross-Embodiment Results Make Sense

On an unseen task, if a VLA has never observed a similar motion in robot data, it can only guess from the semantic prior of its VLM. But a VLM has not been pretrained on action sequences at scale; it does not know how physical motions such as untying shoelaces, ironing clothes, or shaking hands should unfold.

WAM is different.

It can first use its video-generation backbone to produce a future visual plan. That backbone may already have seen similar motion patterns in large-scale video, after which the IDM converts the visual future into robot actions.

There is nothing mysterious about WAM outperforming VLA on unseen tasks.

The same logic applies to cross-embodiment transfer. If a video-only demonstration from a human or another robot can update WAM’s representation of how the task should visually unfold, it may not need action labels. Action labels are embodiment-specific, while future visual states are closer to embodiment-agnostic.

This is WAM’s most valuable property: it turns large amounts of video without robot-action labels into training signals that can shape a policy.

That is difficult for a conventional VLA to do directly.

But the Cost Is High

WAM’s strength comes from explicit world modeling. Explicitly generating a future video, however, is not free.

The model must truly understand the additional information in how the world evolves, which consumes model capacity.

Table 4 makes this clear:

The 5B DreamZero reaches only 21% task progress, while the 14B model reaches 50%.

A small model struggles to carry explicit video generation and action alignment at the same time. It hallucinates the future, and the action fails with it.

WAM is therefore a large-model system that embeds a world model inside the policy. Its engineering profile differs from many VLAs: a VLA can be cheaper, faster, and easier to deploy, while WAM moves much more capability into the video-generation backbone.

The Awkward Part Is Industrial Deployment

WAM looks excellent in zero-shot and low-shot settings. But if the goal is deployment in a factory, warehouse, restaurant, or home-service scenario, one unavoidable issue remains:

You still have to post-train.

Figure 10 is worth examining closely.

On shirt folding, after 33 hours of post-training data, DreamZero and the pretrained VLA are tied at 92.5%. On fruit packing, after 12 hours, DreamZero reaches 96% and clearly outperforms the VLA.

This creates an awkward tradeoff:

With very little data, WAM is stronger than VLA, but it may still not be reliable enough to deploy. As the amount of data grows, the gap between WAM and a strong VLA may narrow, while WAM retains substantially higher inference cost, latency, and system complexity.

The paper acknowledges that DreamZero needs an entire stack of system optimizations to reduce a 14B autoregressive video-diffusion model from 5.7 seconds to 150 ms, and it uses 2×GB200 GPUs to reach 7 Hz. A single GB200 costs more than RMB 500,000 and requires a direct-liquid-cooling (DLC) system. Meanwhile, current VLAs can run at 20 Hz+ on consumer GPUs.

This is not a small engineering gap. It is a deployment-cost gap.

If an industrial task can already support the collection of dozens of hours of high-quality data, a VLA may remain an exceptionally strong baseline because it is cheap, fast, and stable.

WAM’s advantages are more likely to appear in:

• long-tail tasks;
• generalization to unseen motions and low-repetition demonstrations;
• cross-embodiment transfer; and
• the use of video-only data.

It is not a direct replacement for VLA in every scenario.

My Take

WAM is a highly important direction. Its core contribution is not that “robots can generate videos,” but that it demonstrates:

A video-generation backbone can serve as a physical prior for a robot policy.

This direction is more closely aligned with how the physical world evolves than a conventional VLA, and it can make better use of video at scale. It changes robot learning from “directly fit the action” to “understand the future first, then infer the action.”

That is powerful. But it is too early to declare VLA dead.

VLA’s weakness is physical generalization; WAM’s weaknesses are cost, latency, capacity requirements, and deployment complexity.

A more realistic assessment is:

WAM will raise the ceiling on unseen tasks, low-shot learning, cross-embodiment transfer, and video-data scaling; but in many industrial settings, VLA will continue to thrive because it is cheap, fast, and easy to deploy.

WAM exposes the ceiling of VLA. It does not yet prove that VLA’s engineering value has disappeared.

VLA is not dead. But from now on, robot foundation models cannot tell their whole story through VLM pretraining alone.