World Action Models Explained: How WAMs Differ From VLAs in Robotics

A world action model (WAM) is a robot policy built on a pretrained video model that predicts future scenes and robot actions together. The reference result is NVIDIA's DreamZero (arXiv 2602.15922), which reports 62.2% average task progress against 27.4% for the best pretrained vision-language-action (VLA) baseline in unseen environments. Those are the authors' numbers, from 80 real-robot rollouts per checkpoint, so treat them as a strong signal and not a settled result.

Reading note: the DreamZero analysis below is based on the full paper text. The other papers (Cosmos Policy, Fast-WAM, GE-Act 2.0, the WAM-versus-VLA generalization study and the survey) were read at abstract level only, so only their headline claims are used.

What is a world action model?

A WAM is a robotics model that learns how the world is likely to change and which actions shape that change. NVIDIA's glossary defines it as a model that jointly predicts future world states and robot actions using video pretraining.

A September 2026 survey describes WAMs as models that couple future world prediction with executable action generation. Its overview describes two routes: predict futures and actions together, or predict a future and infer the actions from it with an inverse dynamics model.

NVIDIA's June 2026 technical blog splits WAMs into three formulations:

• Inverse dynamics: generate future video, then infer the action.
• Joint prediction: emit video and actions in one model (DreamZero's approach).
• Representation-only: use the video backbone for training but skip video generation at inference.

How is a WAM different from a VLA?

A VLA starts from a vision-language model (VLM) and adds an action module. A WAM starts from a video model. NVIDIA argues that a VLM backbone learns to describe the world, not predict how it evolves, which is the gap a video backbone is meant to fill.

The same NVIDIA blog calls this a second major recipe for robot foundation models, alongside VLAs. For the VLA side of the story, see our explainer on RT-2.

How does DreamZero work?

DreamZero is a 14B autoregressive video diffusion model built on a pretrained video diffusion backbone. It denoises future video and actions jointly, so the predicted frames and the motor commands come from one model.

The paper writes the policy as a video-prediction term multiplied by an inverse-dynamics term, which is why the authors say actions are tied to what the model imagines. They also report that most failures come from video generation errors and not from action prediction. In their words, the policy faithfully executes whatever trajectory the video predicts.

The authors pretrained separately for two robots: the AgiBot G1 bimanual platform (about 500 hours of teleoperation data they collected) and a Franka arm using the DROID dataset. They kept compute comparable across methods by matching batch size and gradient steps.

What are the results?

On the authors' real-robot tests, DreamZero beats pretrained VLAs on every headline comparison. The table lists the lab-reported numbers with evaluation size, all from the full paper.

Two ablations point the same way, though both used only the easy pick-and-place subset and a reduced training budget (50K steps, batch size 32). Diverse data lifted task progress from 33% to 50% against repetitive data of equal size (500 hours). The 14B model scored 50% against 21% for a 5B model.

The paper also reports cross-embodiment gains. Video-only demonstrations from other robots or humans gave a relative improvement of over 42% on unseen tasks with 10 to 20 minutes of data, and adapting to a new robot took 30 minutes of play data.

Does the evidence hold up?

The direction is plausible, but the sample sizes are small and the authors are also the baseline builders. Four points stand out.

• Small rollouts. The DROID unseen-task test has 2 rollouts per task across 20 tasks, so 40 rollouts. By simple arithmetic, 22.5% is 9 successes, 12.5% is 5 and 7.5% is 3. A handful of runs separates the systems.
• Baselines. The GR00T N1.6-DROID checkpoint was internally trained, and the paper itself notes the pretrained VLAs may have seen some of the unseen tasks during cross-embodiment pretraining. The comparison is useful but not neutral.
• Compute. A naive implementation needed about 5.7 seconds per action chunk. With optimizations, including a few-step variant called DreamZero-Flash, latency fell to 150ms on GB200 hardware, and 7Hz control used 2 GB200 GPUs. The paper notes VLAs run at over 20Hz on consumer GPUs.
• Scope. The authors list sub-centimeter precision tasks as a limitation, say their visual memory is short-horizon (6 seconds), and state they still lack evidence for scaling laws for WAMs.

An NVIDIA blog author, Moritz Reuss, adds that WAMs often trade robot-data efficiency for compute cost, and that robot benchmarking is not solved. His blog is a single source for those two points.

Who else is building WAMs, and what do they report?

Several groups publish WAM variants, though not all start from a pretrained video generator. The headline figures below are lab-reported and were read from abstracts.

NVIDIA's June blog also names Ant Group's LingBot-VA, Rhoda AI's DVA, Sereact's Cortex 2.0 and Mimic Robotics' mimic-video as public video-backbone examples. A separate generalization study reports LingBot-VA at 74.2% on RoboTwin 2.0-Plus and Cosmos-Policy at 82.2% on LIBERO-Plus, while noting pi0.5 can match WAMs on some tasks if trained on diverse robot data. For a LIBERO-Plus record claim, see ProWAM's 85.8%.

On the NVIDIA side, the company says Cosmos 3 comes in 4B, 16B and 64B sizes, and that the 4B Edge policy runs at 15 Hz on Jetson Thor. It also says on its glossary page that Isaac GR00T 2, based on DreamZero research, is slated for release by the end of 2026 and ranks No. 1 on MolmoSpaces and RoboArena (company claims, as of the page dated 18 May 2026).

Is every big lab really chasing WAMs?

No. WAMs have real momentum, but not consensus. NVIDIA's own analysis says the field has two major representation bets, and that many teams build on the VLM-based VLA recipe, including NVIDIA GR00T, Xiaomi Robotics and Being-H0.5.

The loudest advocate is NVIDIA's Jim Fan. According to Eventual's summary of his Robotics' End Game talk at Sequoia's AI Ascent 2026, he argues VLAs are being replaced by WAMs. That is a paraphrase of a secondary summary, and Fan works at NVIDIA, which builds DreamZero.

Reuss is more cautious. He writes that WAMs are still in an exploration phase and that the likely next generation is WAM and VLA hybrids. For publication patterns around one VLA lab, see Physical Intelligence's research gap since pi0.7.

What does a working researcher take from this?

Three practical points follow from the evidence so far.

1. Check what you are paying for. If your task needs broad generalization across scenes and your robot data is diverse, the DreamZero ablation (33% to 50%) suggests data diversity matters for WAMs. If you need 20Hz control on a consumer GPU, a VLA is still the easier fit per the paper's own numbers.
2. Test the cheap variant. Fast-WAM's abstract reports that removing video co-training hurt much more than removing test-time imagination. If that holds, you may get most of the benefit without paying for video generation at run time. See also the Tsinghua result on first-denoising-step WAMs.
3. Read eval sizes before ranking. A 2-rollouts-per-task protocol cannot separate close systems. Memory is one open weakness; see NVIDIA's Long-WAM and Video Prediction Policy 2 for follow-up work.

What's still unclear?

• Is video generation at inference needed? DreamZero generates video and attributes failures to video errors. Fast-WAM suggests the value may sit in training. Reuss says the simulation evidence is not enough to settle it.
• Does the win come from the video prior or from the data recipe? DreamZero changes the backbone and the data mix (diverse, non-repetitive) at once. The ablations isolate data diversity and model size, but only on easy pick-and-place tasks.
• Does it scale? The DreamZero authors say they lack evidence for scaling laws for WAMs. AgiBot's GE-Act 2.0 reports gains from 300 to 30,000 hours of data, but that is a different model trained from scratch, by a different group.
• Is a pretrained video generator required? GE-Act 2.0 initializes its generative and action components from scratch on manipulation data, so the "video prior" story may not be the only route.
• Independent replication. The generalization study evaluates released WAMs on simulation benchmarks. The generalization study does not re-run DreamZero's real-robot comparisons.
• Which numbers are comparable? Task progress, success rate and benchmark averages differ across papers, so the table above should not be read as a leaderboard.