Three scenarios for 2027 to 2030

Why this chapter exists

A book about a field that changes monthly should not end with a prediction. It should end by giving the reader a way to recognise which future is arriving. This chapter turns the open problems of chapter 17 and the bets of chapter 18 into three scenarios for 2027 to 2030. It names the evidence that would tell them apart, and closes with a checklist for reading the next model release in about ten minutes.

The scenarios are a device for organising uncertainty, not forecasts. Real outcomes will mix them, and the authors do not pick one.

After this chapter you will be able to:

• describe the three scenarios and the evidence that already points towards each;
• use figure 19.2 to keep score as new results arrive, and say which observations would change your mind;
• read a launch post or paper with the checklist of figure 19.3: place the model on the family grid, find its data on the pyramid, and flag what its evidence does not show.

Scenario 1: consolidation around reactive VLAs

The story. Reactive VLAs absorb the fixes for their weaknesses without changing what they are. Steerable prompts let them learn from failures and select behaviours by description (Physical Intelligence et al. 2026, arXiv 2604.15483). Multi-scale memory carries them through quarter-hour tasks (Torne et al. 2026, arXiv 2603.03596). RL from deployment makes them faster and more reliable than their demonstrators (Physical Intelligence et al. 2025, arXiv 2511.14759; Xu et al. 2026, arXiv 2604.23073). Embodied reasoning models above them handle planning and success detection (chapter 10). World-action models stay a research niche for contact-rich tasks, where predicting consequences is worth its cost.

Why it is plausible. The family has the largest open ecosystem, the cheapest inference and the longest deployment record (chapters 5 and 15). The March 2026 robustness study found that VLAs such as π0.5 can match world-action models on some tasks when trained extensively on diverse data (Zhang et al. 2026, arXiv 2603.22078).

What to watch. Independent real-robot studies at matched data that find no robustness gap between the families. VLA inference staying cheap on edge hardware while video-generating models do not get cheaper.

Scenario 2: convergence on world-action models

The story. Predicting the future becomes a standard part of every deployed policy, mostly through designs that cost little at inference: train with video, act without generating it. The words "VLA" and "WAM" stop meaning different things, because every serious model is trained on both actions and futures.

Why it is plausible. The same robustness study found world-action models strong under perturbation, with LingBot-VA at 74.2 percent on RoboTwin 2.0-Plus and Cosmos Policy at 82.2 percent on LIBERO-Plus (Zhang et al. 2026, arXiv 2603.22078). On RoboArena a world-action model led the VLAs it was compared with (chapter 6, figure 6.7)leaderboard claim (Atreya et al. 2025, arXiv 2506.18123). The cost argument is being dismantled from several sides: Fast-WAM drops generation at test time, Privileged Foresight Distillation keeps the future's benefit without the future, and LingBot-VLA 2.0 adds future prediction to a VLA as a training-only task (Yuan et al. 2026, arXiv 2603.16666; Fang et al. 2026, arXiv 2604.25859; Wu et al. 2026, arXiv 2607.06403). NVIDIA previewed GR00T N2 as a world-action model for humanoidsvendor claim (NVIDIA Newsroom 2026).

What to watch. The robustness advantage replicating outside the labs that reported it. A GR00T N2-class model shipping at control rates on robot hardware. Training with video becoming the default in new releases from the VLA side.

Scenario 3: resurgence of latent and geometric prediction

The story. Generating pixels turns out to be the wrong target for control. Policies predict in learned feature spaces or explicit geometry, pretrain on latent actions from human video, and keep RGB video generators only as simulators and data engines (chapter 8).

Why it is plausible. Latent world-action models already run far faster than pixel ones: LaWAM reports up to 24 times lower latency (Chen et al. 2026, arXiv 2606.15768). JEPA-style objectives avoid spending capacity on what cannot be predicted (Sun et al. 2026, arXiv 2602.10098; Lin et al. 2026, arXiv 2608.09381). Geometry makes the variables of contact explicit, and touch shows what cameras miss (Huang et al. 2026, arXiv 2601.03782; Higuera et al. 2026, arXiv 2602.06001). Even the test-time scaling work on pixel world-action models selects among rollouts using geometric consistency (Zhao et al. 2026, arXiv 2607.17454).

What to watch. Matched-compute comparisons that favour latent targets over pixel prediction. A shared benchmark for geometry-first models. Latent-action pretraining on human video beating teleoperation-only pretraining in an independent study.

What would tell them apart

Each scenario makes predictions the others do not. Figure 19.2 lists ten observable results and what each would mean for each scenario. Tick the ones you have seen reported, preferably by someone other than the model's maker, and watch the scores.

Two things stand out in the table. Most of the discriminating results are evaluations, not new models: independent replications, matched-compute comparisons, shared benchmarks. Chapter 17 rated evaluation as severe and under-researched, and the scenarios inherit that. The future of the field will be decided partly by who builds the experiments that compare its families fairly. Second, several results move two scenarios at once in opposite directions, which is why one result can reverse the lead. Exercise 19.2 asks what it takes to be ahead.

How to read the next model release

New models will arrive faster than this book can be revised. The tools of the earlier chapters make a ten-minute reading possible. Figure 19.3 is the worksheet.

The checklist in words:

1. Family. What does it predict: actions only, actions and futures, futures only, or a plan in language? In what space: pixels, learned features, geometry, or tokens? That places it in chapters 5 to 11.
2. Data. Which layers of the pyramid does it learn from, how many hours of each, and are those hours measured the same way (chapter 13)?
3. Release. Are the weights released, under what licence, and does the backbone carry its own terms (chapter 5)?
4. Footprint. How large is it, and what hardware does it need?
5. Latency. What control rate does it reach, on which hardware, synchronous or asynchronous (chapter 15)?
6. Evaluation. Real robots or simulation? Which benchmark, and is it saturated?
7. Trials. How many per condition, with what intervals (chapter 11)?
8. Who measured. The maker, or someone else?
9. Claims. Are numbers about the maker's own data and model presented as claims?
10. Comparison and failures. Are baselines matched for data and compute, and are failure cases shown?

Several flags are normal, and are a reason to wait for evidence rather than to dismiss the model. What the checklist protects against is the commonest mistake in reading this field: treating a maker's demonstration as a measurement.

In code

The crate's evidence table is the data behind figure 19.2, with the release questions as a struct.

/// What a launch post or paper says about a new model, answered yes or no by the reader.
#[derive(Clone, Copy, Debug, Default)]
pub struct Release {
    pub weights_and_license_stated: bool,
    pub size_and_hardware_stated: bool,
    pub latency_on_named_hardware: bool,
    pub real_robot_evaluation: bool,
    pub trials_per_condition: Option<u32>,
    pub intervals_reported: bool,
    pub independent_evaluation: bool,
    pub own_numbers_flagged: bool,
    pub baselines_matched: bool,
    pub failures_reported: bool,
}

Exercises

The stubs are in rust/ch19-scenarios/src/exercises.rs. This chapter is equation-free; the exercises are bookkeeping for judgement.

cargo test -p ch19-scenarios --test exercises ex19_1

cargo test -p ch19-scenarios --test exercises ex19_2

cargo test -p ch19-scenarios --test exercises ex19_3

Further reading

• The WAM survey's section on open challenges and future directions (Lu et al. 2026, arXiv 2609.16074).
• The robustness study comparing world-action models and VLAs, the closest thing to a scenario test so far (Zhang et al. 2026, arXiv 2603.22078).
• Moritz Reuss's view on the second recipe for robot foundation models, quoted in chapter 6 (NVIDIA Technical Blog 2026).
• Rodney Brooks' annual predictions scorecard, for how to keep yourself honest about the future (Rodney Brooks 2026).
• The industrial readiness survey, for what deployment will demand whichever scenario wins (Kube et al. 2026, arXiv 2603.06749).