Frontier Questions and Research Roadmaps › Limitations and Impact

The three experiments are hand-constructed and low-dimensional. Experiment 1 uses a known stable discrete transition law with a small drag term and an autonomous nuisance process. Experiment 2 uses a known transition family with a known coefficient. Experiment 3 uses two known modes and a heuristic assumed hazard, tested against stable and deterministic-change conditions. Known truth permits controlled comparisons, but the numerical performance results do not establish physical-system performance. In experiment 1 the aggregate metric changes yet orders two models oppositely to controller cost; it does not establish that decision consequence by itself. The examples demonstrate scoped failure mechanisms, not the magnitude of their effects on real systems.

These examples motivate requirements for the proposed studies, not a diagnosis of what limits the whole field:

• General-purpose worlds need specified action semantics and held-out evaluation populations for the transfer being claimed. A single canonical interface is one design option, not a necessary architecture. Open X-Embodiment provides bounded policy-transfer evidence across its specified robot datasets; dynamics-model transfer needs its own evaluation.
• Causal, compositional, persistent structure calls for appropriate identification assumptions, informative tests and change-aware update rules. Model capacity, optimization, data coverage and experimental design can all matter. Experiment 2 supplies one tractable combination split; project 3 proposes an interventional comparison whose implementation cost must be measured.
• Evaluation beyond visual prediction needs measurements appropriate to the decisions being claimed. One scalar can summarize a specified objective, but it does not establish every aspect of usefulness or safety. Experiment 1 separates aggregate observation error from closed-loop cost; report coordinate-wise and control measurements to explain that discrepancy.

An unspecified universal-model claim leaves the evaluation population and criterion unclear. A transfer contract makes comparisons interpretable. Identification, calibration, change detection and fallback design deserve explicit testing alongside capability development; an impressive demonstration alone does not establish deployment reliability.

The methods in Part VIII use learned models in different ways. PETS combines probabilistic ensembles with trajectory sampling for control. MBPO trains with short model-generated rollouts branched from real data. DreamerV3 improves behavior using imagined futures. MuZero uses tree search with a model predicting planning-relevant rewards, policies and values. TD-MPC2 performs local trajectory optimization in a learned latent model. These are not all instances of policy training in imagination. Their different model interfaces also require different validation tests.

Verification deserves explicit effort alongside capability development. These experiments do not establish the field's binding constraint or the feasibility of every proposed project for one researcher. They provide runnable examples and study specifications that can be reduced to pilot experiments. A useful pilot measures its actual computational cost, checks its identification assumptions and estimates what effect size its evidence can resolve.