World action models vs VLAs: predict the video, or just the moves

VLA and WAM are the two names you keep hearing in robot learning this year. Both are robot policies: they look at a camera image, read an instruction like "put the bowl on the plate," and output the robot's next moves. So what's different? Mostly the model they start from and what they practice predicting.

Same job, different start

A vision-language-action model (VLA) starts from a vision-language model, trained on images paired with text. RT-2 named this category in 2023, and π0 is a well-known one. Most recent world action models (WAMs) start from a video generation model instead, trained to continue a clip. Fast-WAM takes the name from DreamZero (February 2026), though the words are older. On the robot, both answer the same request, the current image $o_t$ and instruction $\ell$ in, a chunk of the next $H$ moves out:

$$a_{t:t+H} \sim \pi_\theta(\,\cdot \mid o_t,\ \ell\,)$$

WAMs learn one extra target

Take one demonstration: the frame now, the frames the camera saw next, and the moves the robot made. A VLA learns the moves, and some VLAs also keep training on image-text tasks. A WAM learns the moves and the next frames $o'$ together:

$$\text{VLA: } p_\theta(a \mid o, \ell) \qquad\qquad \text{WAM: } p_\theta(o', a \mid o, \ell)$$

In Fast-WAM that's one loss for each part, $\mathcal L = \mathcal L_{\text{act}} + \lambda\,\mathcal L_{\text{vid}}$. The bet, which Fast-WAM takes from DreamZero and LingBot-VA (Section 2), is that a video model already knows how scenes change, and image-text pretraining doesn't model that directly.

Many WAMs imagine, then act

Many WAMs, early and recent, draw the future while the robot runs. Some denoise video and moves together. Others generate the video first, then read the moves off it with an inverse dynamics model, which infers the action between two frames. Either way, each chunk runs a big video network many times:

$$T_{\text{chunk}} \approx S_v\,c_{\text{video}} + S_a\,c_{\text{act}}$$

$S$ counts denoising steps and $c$ is the cost of one pass. LingBot-VA on RoboTwin uses $S_v = 25$ and $S_a = 50$ (Faster-WAM, Appendix E.3).

Fast-WAM imagines only in training

Fast-WAM (March 2026) keeps the video loss in training and drops the imagined frames on the robot. The video model makes one pass over the current frame, then an action expert turns noise into 32 moves in 10 steps:

$$T_{\text{chunk}} \approx c_{\text{video}} + 10\,c_{\text{act}}$$

On the authors' towel-folding robot, with one RTX 5090D GPU, that's 190 ms per chunk, against 580 and 810 ms for two versions of itself that imagine. Speed isn't the whole story there. The VLA π0.5 took 180 ms and had the best success, and the frames-first version beat Fast-WAM on success. A π0.5 without its robot pretraining did worse than every Fast-WAM version trained with the video loss.

What the video loss buys

Did dropping the imagined frames cost anything? The same paper trains four versions on one backbone. Taking away the video loss costs 4.1 points on LIBERO and 8.0 on RoboTwin. The two imagining versions land within 1.2 points of Fast-WAM, a bit above it on LIBERO and below it on RoboTwin. So on these two simulated benchmarks, the video loss in training matters much more than drawing frames on the robot. Each bar is one run, so gaps near one point can be noise. All four keep the video-pretrained trunk, so this test doesn't measure what the trunk itself adds.

Under distribution shift, the gap gets much bigger. The other Faster-WAM paper trains Fast-WAM and a version that denoises frames and moves together with the same backbone, data and recipe. On LIBERO-Plus, Fast-WAM scores 49.1% and the joint version 66.3%. On a real robot with new backgrounds, lighting and distractors, Fast-WAM gets 45.6% and the joint version 55.6%. But the Faster-WAM in the next section reaches 78.3% on LIBERO-Plus without any test-time frames, so why Fast-WAM drops is still an open question.

Score against speed

Here are both families on one chart, from the RoboTwin 2.0 table in the Faster-WAM paper. Motus and LingBot-VA, which imagine on the robot, score 87.8% and 92.2% but take 1.4 and 7.4 seconds per chunk. The VLAs π0 and π0.5 answer in 65 and 71 ms with 62.2% and 79.8%. Fast-WAM keeps the WAM score at about 200 ms, and Faster-WAM gets to 53 ms at 90.5% with a one-layer action expert that reads from every video layer. Faster-WAM in LeRobot shows how.

It isn't a clean race, though. The four baselines were pretrained on robot data, and Fast-WAM and Faster-WAM weren't. π0 and π0.5 predict 50 moves per chunk and Faster-WAM 32, so per move π0.5 is slightly faster. And on LIBERO-Plus, which perturbs the camera, lighting, language and more, π0.5 leads with 85.7% against 78.3% for Faster-WAM.

Three WAM branches, not one line

It's tempting to read this as one line from slow WAMs to fast ones, but it isn't. VPP (2024) and UVA (2025) already acted from video-model features without drawing frames, and GR-1 learned future images and actions together in 2023. The lanes follow the Fast-WAM paper's grouping, and GR-1's place follows the June survey. The Faster-WAM paper sorts by whether video and actions share one network, so LingBot-VA lands somewhere else.

Two August 2026 papers are called Faster-WAM. arXiv 2608.02365 builds on Fast-WAM and cuts the action expert to one layer. arXiv 2608.04404 keeps a future representation, computes it once, and reuses it while it denoises the actions.

A label that took off

On arXiv, papers with "world action model" in the title or abstract went from 12 in the first quarter of 2026 to 81, then 214. VLA papers grew from 357 to 646 over the same quarters, so the ratio went from about 1 in 30 to 1 in 3. A June 2026 survey says the label "has spread faster than the understanding behind it." The count misses WAMs that use other words, like LingBot-VA, and catches some driving papers, so it tracks a word more than a method.

So where is the line?

On the robot, Fast-WAM and Faster-WAM look just like a VLA: image and instruction in, moves out. What makes them WAMs is the video trunk and the video loss that trained it. The June survey asks whether the predicted future helps produce, score or train the actions inside one model, and it counts Fast-WAM because its action path "still uses a representation shaped by future prediction." It also excludes "a direct VLA with an auxiliary future loss", so part of the line is a matter of definition.

Symbols, losses and the time formula

The policy. $o_t$ is the camera image at time $t$ (some setups add a wrist camera and the robot's joint readings, and some WAMs, like LingBot-VA, also read a history of past frames and actions), $\ell$ is the instruction, and $a_{t:t+H}$ is the chunk of the next $H$ actions. Fast-WAM uses $H = 32$. $\pi_\theta$ is the policy with weights $\theta$, and $\sim$ means the chunk is sampled, because a flow-matching head turns random noise into a chunk.

The losses. Both losses here are flow-matching losses. During training, the network sees a mix of the target and random noise and learns the velocity between the two. On the robot, a sampler follows that velocity from noise to a clean chunk. $\mathcal L_{\text{act}}$ does this for the demonstrated action chunk. $\mathcal L_{\text{vid}}$ does it for the future frames, after the video model's autoencoder has compressed them into latents. $\lambda$ sets the weight of the video term. The Fast-WAM paper doesn't give its value, and the Faster-WAM paper sets both weights to 1. π0 trains its action expert with the same kind of action loss, so the action head can be the same in both families. The trunk and the extra target are what differ.

The joint form. $p_\theta(o', a \mid o, \ell)$ fits the WAMs that denoise frames and moves together. Fast-WAM blocks the action tokens from seeing the future video tokens, so it really learns $p_\theta(a \mid o, \ell)$ and $p_\theta(o' \mid o, \ell)$ with one shared trunk.

The time formula. $T_{\text{chunk}} \approx S_v\,c_{\text{video}} + S_a\,c_{\text{act}}$ is a back-of-the-envelope model, not a measurement. It ignores the image and text encoders and assumes each denoising step costs one full pass. It's there to show why $S_v$ matters when the video network is the big one.

Where the definitions disagree

• Who coined "world action model". DreamZero writes "We term this architecture a World Action Model (WAM)", and Fast-WAM says it follows DreamZero. The words are older. JOWA (2024) used them for offline reinforcement learning on Atari games, DyWA (March 2025) for a manipulation policy that predicts future states with its actions, and Genie Envisioner (August 2025) called its policy GE-Act "a world action model". "Vision-language-action" also shows up in an arXiv title before RT-2, but RT-2 is the paper that named the category.
• Does every WAM start from a video model? Most recent ones do, including DreamZero, Fast-WAM and Faster-WAM. The June survey also counts a parallel line built on language or vision-language backbones, as long as a predicted future feeds the actions.
• Is Fast-WAM a WAM at test time? Its authors say its test-time interface is "similar to standard VLA policies". The June survey still counts it, because its action path uses a representation shaped by future prediction.
• Which group is LingBot-VA in? Fast-WAM puts it under "video first". Its own abstract describes a Mixture-of-Transformers. Faster-WAM calls it shared-backbone, based on its released code.

Timing details for the speed chart

• Who timed what. The Faster-WAM authors timed every model themselves on one 24 GiB consumer GPU, with batch size one, counting model compute only. The success scores of the other models come from their own papers.
• Software. π0 and π0.5 ran as compiled JAX code. The WAMs ran in PyTorch without torch.compile.
• Memory. Fast-WAM, Motus and LingBot-VA don't fit on that GPU as complete pipelines, so their times add up parts that were timed one at a time.
• Moves per chunk. π0 and π0.5 predict 50, Motus 16, and LingBot-VA, Fast-WAM and Faster-WAM 32. Per move, π0.5 takes about 1.4 ms and Faster-WAM about 1.7 ms.
• π0.5 in two papers. π0.5 shows 71 ms here and 180 ms in Fast-WAM's real-robot timing. Neither paper explains the gap. The GPU doesn't explain it either, because Fast-WAM takes about 190 ms in both papers.

Sources for each drawing

1. RT-2 2307.15818, π0 2410.24164, DreamZero 2602.15922 Section 1, Fast-WAM 2603.16666 Section 2, Faster-WAM 2608.02365v2 Section 1.
2. Fast-WAM Section 3.2 (joint flow-matching objective). Faster-WAM Appendix C (both loss weights 1).
3. Fast-WAM Section 1 (imagine-then-execute, the two groups and their examples). Faster-WAM Appendix E.3 (LingBot-VA steps).
4. Fast-WAM Sections 3.2 and 4.1 (one pass, 32 actions, 10 steps) and Figure 4 (latency, read from the PDF page). Section 4.3.3 (π0.5 strongest on towel folding, the frames-first version highest among the Fast-WAM versions, π0.5 without pretraining below every version trained with the video loss).
5. Fast-WAM Tables 1 and 2 (average success on RoboTwin and LIBERO) and Section 3.3 (the version without the video loss keeps the architecture). The other Faster-WAM 2608.04404, Experiments (controlled setup, LIBERO-Plus and real-robot shifts).
6. Faster-WAM Table 2 (RoboTwin 2.0), Table 3 (LIBERO-Plus), Appendix E.1 and E.2 (latency setup) and Table 8 (moves per chunk).
7. v1 dates on arXiv: UniPi 2302.00111, RT-2 2307.15818, GR-1 2312.13139, OpenVLA 2406.09246, π0 2410.24164, VPP 2412.14803, UVA 2503.00200, UWM 2504.02792, π0.5 2504.16054, Vidar 2507.12898, Motus 2512.13030, LingBot-VA 2601.21998, DreamZero 2602.15922, Fast-WAM 2603.16666, Faster-WAM 2608.02365 and the other Faster-WAM 2608.04404. Lanes follow Fast-WAM Sections 1 and 2.
8. arXiv API, run on October 22 and again on October 23, 2026, with the same totals for every full quarter: ti:"PHRASE" OR abs:"PHRASE" with a submittedDate range for each quarter, where PHRASE is world action model or vision language action. The API stems words, so the plural matches too. A paper with both phrases counts in both series. Dates are first-version dates, so a paper renamed later counts in its first quarter: 2602.15882 was FUTURE-VLA in version 1 and PACT-WAM in version 2. October 1 to 22 gives 104 WAM papers and 262 VLA papers. Survey quote: 2606.20781, Section 1.