Towards Generalist Embodied AI: A Survey on World Models for VLA Agents
1. Paper overview
In one sentence: The survey classifies world models by how predicted futures enter an agent’s workflow, while its compiled scores leave causal gains and real-world reliability unresolved. e02e03e04e10e11e13e16
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors identify a gap between semantic competence and physically executable behavior. Text-centric reasoning can propose plans without verifying continuous dynamics, while robot demonstrations are scarce and physical RL is costly or risky. Their organizing question is how future predictors support generalist VLA agents across planning, action generation, training data, and feedback. e02 |
| Core mechanism | |
| A key reported result | LIBERO manipulation across Spatial, Object, Goal, and Long suites: SRPO (Online): Spatial 98.8%, Object 100.0%, Goal 99.4%, Long 98.6%; Avg. 99.2%. Success rate (%) and reported suite average. Survey compilation of simulation success rates; online and offline rows are explicitly distinct. A common data, seed, trial-count, or compute protocol is not supplied. RynnVLA-002: Avg. 97.4%; SRPO (Offline): Avg. 92.5%, Long 88.7%. SRPO (Online) has the highest listed average. This cross-method comparison is not a controlled ablation of online interaction or world-model architecture. e10e16 |
| Reading caution |
Core contributions
Figure 2. One system can occupy several roles in the survey’s historical map. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across each colored row to follow a role, then look vertically for the same system appearing elsewhere. The rows distinguish world planners, world action models, world simulators, and world synthesizers. GR-MG appears in both the planner and action-model rows. Genie Envisioner appears in the planner, simulator, and synthesizer rows. Those repeated names matter: a contribution may be deployed in different parts of a learning pipeline. Also inspect the time labels before interpreting density. The first two columns cover whole years, whereas the columns within 2025 cover quarters, so equal horizontal widths do not represent equal durations. e03
What it supports. The authors portray a widening field, with planners and action models preceding the plotted synthesizer and simulator activity. Reader interpretation: the repeated entries make the taxonomy most useful as a map of system roles. Counting each name as one mutually exclusive paper category would lose information already visible in the figure.
Where the evidence stops. The figure is a selected literature timeline, not a normalized publication-rate measurement or a systematic-search audit. Its dates and category placements are the survey’s account; the underlying papers were not independently checked in this reading.
2. Motivation
2.1 The problem and the proposed response
The authors identify a gap between semantic competence and physically executable behavior. Text-centric reasoning can propose plans without verifying continuous dynamics, while robot demonstrations are scarce and physical RL is costly or risky. Their organizing question is how future predictors support generalist VLA agents across planning, action generation, training data, and feedback. e02
2.2 What this reading follows
A future predictor can help a robot in several different ways. It can supply a policy with imagined guidance, generate actions together with observations, manufacture demonstrations, or provide an environment for evaluating and improving a policy. This survey turns those uses into four organizing paradigms and connects them to model backbones, metrics, and benchmarks. Read the diagrams as descriptions of information flow, then examine what the score tables actually measure. The strongest reported simulation averages are impressive, but the survey does not put all methods through a common experiment. The distinction between a useful taxonomy and verified physical competence is central to this reading. e02e03e04e10e11e13e16
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Related resources. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational survey/resource classification is supported. Architecture, prediction paradigm, and quadrant are not applicable to this survey. Its conceptual joint-model category does not establish that every listed implementation satisfies a One Model architecture criterion; that requires inspecting each model. e03e04e06e16
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Classify a use of prediction before classifying a model
Begin with the destination of a prediction signal. If a future image or latent conditions a separate policy, the survey calls that use a world planner. If future observations and actions are modeled jointly, it is a world action model. If generated trajectories become imitation data, it is a world synthesizer. If an action-conditioned predictor provides an environment for evaluation or reward-driven improvement, it is a world simulator. This reading follows Figure 3 and its associated definitions. My interpretation is that the most informative unit of classification is the system’s use in a pipeline, rather than a paper title alone. Figure 2 supports that interpretation by repeating Genie Envisioner across three roles. The survey’s conceptual joint-output block should therefore not be used as sufficient evidence of a particular implementation’s parameter sharing. e03e04e05e06e07e08
Figure 3. Follow where the prediction goes: guidance, joint output, imitation data, or reward feedback. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. In panel (a), the current observation feeds the world model and VLA; the future signal z links the world model to the policy. Panel (b) groups future observation and action prediction inside one conceptual world action model. In panel (c), follow the loop between generated observations and policy actions, then the synthetic dataset’s path into imitation learning, marked IL. Panel (d) routes an imagined trajectory to the external evaluator Rext, whose feedback returns through RL. Equations (1)–(4) and Sections 3.1–3.4 support these directions. The section introduction says that history and language conditioning are omitted from the notation. e04e05e06e07e08
What it supports. The diagram distinguishes what future prediction is used for. Training a policy on generated trajectories is a different use from conditioning an action on an imagined future or improving it using imagined rewards. This explains why a shared backbone can support several roles without those roles implying the same training or deployment pipeline.
Where the evidence stops. The panels are abstractions, not implementation-level architectures. Panel (c) illustrates action-conditioned synthesis; Section 3.3 also describes action-free video generation followed by inverse dynamics. Neither joint outputs nor the drawn feedback loops certify physical correctness.
5.2 Trace how images become actions, data, or feedback
The planner branch separates predicting a future from extracting control. Section 3.1 describes pixel-level futures processed by inverse dynamics, latent-space prediction, and video-diffusion embeddings that guide policies. The synthesizer branch reuses related ingredients for a different purpose: an action-conditioned rollout interleaves policy actions and imagined observations, while an action-free generator needs inverse dynamics to attach actions to visual trajectories. Those trajectories support imitation learning. The simulator branch instead scores imagined outcomes with an external evaluator, with evaluation-only, reinforcement-learning, and test-time-adaptation variants. These distinctions explain why training and inference must be traced separately. A world model can contribute training data without being called by the deployed policy; conversely, the survey describes VLA-Reasoner invoking MCTS and AdaPower using test-time training. Neither arrangement alone demonstrates reliable physical execution. e05e07e08e13
5.3 Read the strongest score together with its evidence boundary
For LIBERO, start with SRPO (Online)’s 99.2% average, then retain its setting label and suite scores. For CALVIN, distinguish DreamVLA’s average chain length of 4.44 from its 78.1% five-instruction completion rate; TriVLA leads that final column at 81.8%. These are different summaries of behavior. The authors use the leading averages to argue for more demanding benchmarks. My reading is narrower: the compilation documents high reported simulation performance, but it does not harmonize training data, online interaction, compute, or uncertainty. Table 8 helps explain the remaining gap because video quality, flow accuracy, and robot outcomes measure different properties. Table 9 adds another distinction between simulation and real-world data. Testing the survey’s broader physical-reliability claims would therefore require evidence beyond these two leaderboards. e10e11e12e13e14e16
5.4 Training and inference
During training
Equations (1)–(4) summarize policy likelihood under predicted guidance, joint future/action likelihood, synthetic imitation data, and external-reward optimization. The survey specifies no universal stage order, frozen modules, optimizer, training budget, or hardware configuration. e05e06e07e08e16
Table 5 lists Stable Video Diffusion at 1.5B parameters, Emu3 at 8.5B, and V-JEPA 2 at 1B. Reader qualification: these backbone entries do not determine total agent size or comparable training cost. e09
During inference
Planners consume predictions to condition control; action models couple prediction and action generation. Synthesizers can prepare training data without being deployed in the resulting policy. Simulator use includes test-time adaptation: MCTS for VLA-Reasoner and test-time training for AdaPower. e05e06e07e08
Predicted visual plausibility is distinct from executed task success. A control claim needs evidence about action extraction and feedback, beyond the presence of a prediction loss or generated video. e04e12
5.5 Implementation flow
- World planner: predict, then derive control
A world model predicts guidance; a separate policy conditions on it. Section 3.1 covers generated images followed by inverse dynamics, direct latent prediction, and video-diffusion embeddings. Table 1 distinguishes explicit planning with embeddings from implicit representation prediction. e04e05
- World action model: couple vision and control
The generic objective fits future observations and actions jointly. Table 2 distinguishes autoregressive pretraining, unified token streams, foresight and reasoning from discrete and real-valued diffusion. These are literature categories, not one specified architecture or a guarantee of physical consistency. e06
- World synthesizer: turn imagined trajectories into data
Action-conditioned generation alternates policy actions and predicted observations. Action-free generation imagines visual trajectories and infers actions through inverse dynamics. Both support imitation learning; WristWorld is separately classified as view augmentation. Figure 3 illustrates the action-conditioned variant. e04e07
- World simulator: evaluate imagined consequences
An action-conditioned predictor generates outcomes for external reward evaluation. Table 4 separates evaluation-only use, sparse/dense reward RL, and test-time adaptation. Using a world model as an evaluator need not imply a policy update or physical execution. e08
- Match representation and evaluation to the role
Section 4 separates video generators, unified understanding/generation, and representation backbones. Section 5 distinguishes appearance metrics from motion and robot outcomes; Section 6 separates simulation from real-world data and stresses horizon and scene diversity. e09e12e13
6. Experiments & results
This survey organizes world models for vision-language-action agents by what prediction does: guide a policy, jointly generate actions and futures, supply imitation data, or provide an imagined environment for evaluation and improvement. Its benchmark compilation motivates stronger evaluation, but does not isolate the causal benefit of world modeling.
This is a survey with a conceptual information-flow diagram and compiled quantitative comparisons, not a new architecture evaluated through original experiments. It contains no controlled ablation or experimental diagnostic figure, and no common evaluation recipe for Tables 6–7. Six original visuals are included; the metric and benchmark inventories supply evaluation context and are not presented as ablations. Individual cited systems’ implementation claims and performance provenance cannot be resolved from the survey alone. e03e04e10e11e12e13e16
6.1 Read the original evidence
Table 6. The highest listed average is 99.2%, under a compilation that preserves online and offline labels. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the four suite columns before using Avg. as a ranking. Every numerical entry is a success percentage according to the original caption. SRPO (Online) reports 98.8 for Spatial, 100.0 for Object, 99.4 for Goal, and 98.6 for Long, yielding the printed average of 99.2. RynnVLA-002 has the slightly higher Spatial entry, 99.0, but an average of 97.4. The table explicitly separates SRPO (Offline) from SRPO (Online); the former reports 92.5 average and 88.7 on Long. Keep those setting labels attached to the scores when comparing rows or quoting the leading result. e10e14e16
What it supports. SRPO (Online) leads the survey’s average and three of its suite columns. The source uses this result to motivate its near-saturation argument. Reader interpretation: the row establishes the reported simulation score, while the table alone cannot show how much improvement comes from world modeling, extra interaction, data, or other implementation choices.
Where the evidence stops. No uncertainty, seed counts, trial counts, or common training budget accompany these rows. The online/offline comparison is not presented as a controlled ablation, and simulated task success does not establish physical robot performance.
Table 7. Average progress and complete-chain success favor different rows, with an unresolved aggregation inconsistency. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the header: the source identifies CALVIN ABC→D, and the five numbered columns report percentages for consecutively completed instructions. Avg. Len. is measured in completed instructions, not percent. DreamVLA’s row is 98.2, 94.6, 89.5, 83.4, and 78.1, followed by 4.44. UniVLA reports a 4.41 average. TriVLA’s last success column is 81.8, the highest printed five-instruction rate, while its reported average is 4.37. Thus a reader interested in finishing every instruction should inspect the final success column instead of assuming the largest average also wins there. All displayed numbers are preserved from the source. e11e16
What it supports. DreamVLA has the highest listed average, while TriVLA has the highest listed full-chain success. Reader arithmetic also flags a source inconsistency: under a common cumulative-chain interpretation, TriVLA’s five percentages sum to an implied average of 4.410, not the printed 4.37. This calls for checking aggregation provenance before relying on fine ranking differences.
Where the evidence stops. The arithmetic check assumes the columns describe the same chains and weighting. The survey gives no reconciliation or per-chain records. The crop preserves 4.37; 4.410 is a conditional reader calculation, not a corrected experimental result.
Table 8. Match the metric to the claimed capability before treating a better number as better control. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the gray group labels to separate appearance, motion, policy outcomes, and broader benchmark assessments. Within each row, Abbr. gives the short name and Tr. gives the preferred direction: upward triangles mean higher is better, downward triangles lower. Freq. stars indicate usage frequency in surveyed work; they are not performance scores. GT denotes the source’s ground-truth-reliance classification. The top group includes pixel and feature/distribution measures, while the robot-task group asks whether goals or subtasks were completed. The rightmost column supplies example methods using each measure. The lower benchmark group lists broader assessments, rather than another single comparable scalar. e12e13e16
What it supports. The inventory makes a useful evaluation separation: visual fidelity, flow accuracy, and successful action execution answer different questions. Reader interpretation: an image or video quality gain cannot establish a corresponding control gain without task evidence. Likewise, task scores alone leave unanswered whether an imagined world remains geometrically and temporally reliable.
Where the evidence stops. The table’s GT crosses for robot tasks should not be read as eliminating the need to judge task completion. The caption supplies no numerical scale for the frequency stars, and the inventory reports metric choices rather than a correlation study or ablation.
Table 9. Benchmark coverage includes embodiment and scene diversity, not only trajectory volume. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the simulation and real-world groups separately, then move across each row from task domain to data counts. LH. marks the long-horizon setting; S-Arm means single arm. Exo., Ego., and Wrist denote exocentric, egocentric, and wrist-mounted cameras. The final columns count trajectories, scenes, objects, and tasks as recorded by the survey. Dashes are missing entries and must not be treated as zero. BridgeData’s superscript s specifically means skills, so its final count should not be compared directly with ordinary task counts. LIBERO and CALVIN occupy the simulation group, which is the scope of the paper’s two performance tables. e13e14
What it supports. The inventory provides context for the authors’ call for broader evaluation. CALVIN is listed with 24k trajectories across four scenes, while Droid is listed with 76k across 564 scenes. Reader interpretation: trajectory count and scene coverage are different dimensions; neither alone establishes generalization or substitutes for testing the relevant dynamics and task horizon.
Where the evidence stops. These are survey inventory counts, not harmonized training splits or audited current dataset releases. Real-world datasets being listed does not mean the compiled LIBERO/CALVIN scores were measured on physical robots. Missing cells and the skills/task distinction remain unresolved source boundaries.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO manipulation across Spatial, Object, Goal, and Long suites Survey compilation of simulation success rates; online and offline rows are explicitly distinct. A common data, seed, trial-count, or compute protocol is not supplied. | SRPO (Online): Spatial 98.8%, Object 100.0%, Goal 99.4%, Long 98.6%; Avg. 99.2%. Success rate (%) and reported suite average | RynnVLA-002: Avg. 97.4%; SRPO (Offline): Avg. 92.5%, Long 88.7%. SRPO (Online) has the highest listed average. This cross-method comparison is not a controlled ablation of online interaction or world-model architecture. e10e16 |
| CALVIN ABC→D consecutive instruction execution Survey compilation of simulated sequential instruction completion; training/evaluation implementation details are not harmonized in the table. | DreamVLA: Avg. Len. 4.44; successive completion rates 98.2%, 94.6%, 89.5%, 83.4%, 78.1%. Average completed instruction-chain length; success through each of one to five instructions (%) | UniVLA: Avg. Len. 4.41; GR-1: 3.06. TriVLA has the highest listed five-instruction rate, 81.8%, with reported Avg. Len. 4.37. The highest average and full-chain rate belong to different methods. Preserve the reported ranking while flagging TriVLA’s aggregation issue; confidence intervals are absent. e11e16 |
6.3 Ablations and diagnostic examples
No ablation or diagnostic visual is included in this edition. The source-specific visual limitations are recorded above.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors call these simulation benchmarks near-saturated. Reader qualification: high selected scores do not establish uniform saturation, equal training conditions, or safe physical transfer. Compiled tables cannot attribute gains to one mechanism. e10e11e13e14
Physical consistency, 4D perception, safety/reliability, long-horizon foresight, and failure-aware dynamics remain research directions. The authors identify over-optimism from successful-only demonstrations and propose negative or suboptimal experience. These are recommendations, not validated remedies in this survey. e15
Conditional arithmetic check: if CALVIN’s percentages are cumulative completion frequencies for the same chains, TriVLA’s entries imply mean length 4.410, versus the printed 4.37. The source does not explain this discrepancy; no value is silently corrected. e11
7.2 Questions for discussion
- What additional evidence would make imagined success trustworthy for physical execution?
- Can cross-paper rankings survive matched data, interaction budgets, and uncertainty reporting?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproducing the comparison requires method-specific checkpoints, dataset versions, partitions, instruction chains, trial counts, seeds, and online interaction budgets. Tables 6–7 provide no common recipe; backbone parameter counts cannot substitute for it. e09e10e11e16
Proposed checks: audit CALVIN aggregation against matched per-chain logs, and classify representative systems by the role of their prediction signal, allowing multiple roles. These test consistency and taxonomy applicability before implementing a new agent. e03e04e11
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Reconcile the CALVIN row aggregates
Reader-proposed check, not a reproduced experiment: obtain the same-run per-chain completion records behind Table 7, retaining checkpoint, ABC→D split, seed, and weighting. Compute cumulative success at each chain position and mean completed length from those identical records. Use DreamVLA as a positive arithmetic control: its printed percentages imply 4.438, rounding to 4.44. TriVLA’s imply 4.410, versus 4.37. If the original records reconcile the latter only through different weighting or runs, document that protocol difference; if they do not, flag a transcription or aggregation error. Preserve the printed values until provenance resolves which quantity is wrong. e11e16
Check 2: Test whether a single-label taxonomy loses documented roles
Reader-proposed consistency check: have two annotators independently code the entries in Tables 1–4 using the signal destinations in Figure 3, then compare their assignments with Figure 2. Repeat under a forced single-label rule and a rule permitting multiple roles, using the same entries and definitions. Genie Envisioner is a concrete control because the survey places it in planner, synthesizer, and simulator categories; GR-MG tests the planner/action-model boundary. Measure disagreements and the fraction of documented role links retained. If multi-role coding still produces systematic disagreements, identify which definition or pipeline stage remains ambiguous rather than treating a paper name as architectural proof. e03e04e05e06e07e08
8.3 Reading coverage
Visual audit: All ten pages of the supplied v1 were rendered and visually inspected, including cover/manuscript title blocks, Figures 1–3, Tables 1–9, Equations (1)–(4), future directions, conclusion, and references. Six final original crops were separately viewed: Figure 2, Figure 3, and Tables 6–9. Figure 3’s claim-relevant directions were cross-checked against Sections 3.1–3.4 and the equations; panel (c) depicts only the action-conditioned synthesis variant. The CALVIN aggregation discrepancy is preserved and identified as conditional reader analysis. The source contains no original ablation page or appendix. Separate supplements, cited papers, code, and experimental records are outside this inspection.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- 27 January 2026
- Abstract
- 1 Introduction
- 2 Preliminaries
- 3 Taxonomy of World Models for VLA
- 3.1 World Planner
- 3.2 World Action Model
- 3.3 World Synthesizer
- 3.4 World Simulator
- 4 Foundation Models
- 5 Evaluation Metrics
- 6 Benchmarks
- 7 Future Directions
- 8 Conclusion
- References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction-only figure omission was resolved by visually inspecting all ten supplied PDF pages, all three figures, and all nine tables; all six final crops were inspected.
- Identity: the inspected cover and manuscript title page match the catalog title and all nine authors. This report concerns TechRxiv v1 dated 27 January 2026. No different revision was supplied or compared.
- No appendix is present in this ten-page artifact. Separate supplements remain unverified.
- Underlying cited papers and linked code were not inspected; no experiments were run. Compiled scores are verified against this survey only.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF pp. 1–2, title/author blocks; p. 1 footer
The title and nine authors match the catalog: Wentao Tan, Lei Zhu, Bowen Wang, Enci Xie, Baixu Ji, Zengrong Lin, Wenjie Yang, Jingjing Li, and Heng Tao Shen. Affiliations are School of Computer Science and Technology, Tongji University; School of Computer Science and Engineering, UESTC. The cover identifies TechRxiv v1, 27 January 2026, DOI 10.36227/techrxiv.176948355.54623875/v1, and preprint status.
Go to primary source ↓e02PDF p. 2, Abstract, Section 1, Figure 1; p. 3, Section 1 continuation
The survey motivates future prediction for VLA through physical hallucination, unverifiable abstract plans, scarce robot data, and risky on-robot RL. Its scope emphasizes generalist foundation-model-based agents.
Go to primary source ↓e03PDF p. 3, Figure 2 and Section 1, Development Trends and Contributions
The timeline organizes selected works into planner, action model, simulator, and synthesizer roles from 2023 through 2025. GR-MG appears in planner and action-model rows; Genie Envisioner appears in planner, simulator, and synthesizer rows. Contributions comprise primary/secondary taxonomy, ecosystem/evaluation summaries, and future directions.
Go to primary source ↓e04PDF p. 4, Figure 3 and Section 3 introductory paragraph
The panels show future guidance into a VLA; joint observation/action prediction; generated trajectories feeding imitation learning; and imagined trajectories scored by external reward feeding RL. History and text instructions are implicitly conditioned on and omitted from the equations.
Go to primary source ↓e05PDF p. 4, Section 3.1, Equation (1), Table 1, and Evolution
Planners predict future observations or latent guidance for a policy. Pixel-generation approaches can use inverse dynamics; V-JEPA 2 and PIVOT-R are implicit embedding planners, while VPP, MinD, TriVLA, GO-1, and Genie Envisioner are explicit planners with embedding guidance. MoWM is hybrid.
Go to primary source ↓e06PDF p. 4, Section 3.2, Equation (2), Table 2; p. 5, Section 3.2 continuation
World action modeling approximates the joint distribution of future observations and actions. Subcategories include autoregressive video pretraining, unified token streams, foresight, reasoning, discrete diffusion, and real-valued diffusion. WorldVLA and RynnVLA-002 exemplify unified streams; DUST and FLARE exemplify real-valued diffusion.
Go to primary source ↓e07PDF p. 5, Section 3.3, Equation (3), its two factorizations, and Table 3
Synthesizers construct observation-action trajectories for imitation learning. Action-conditioned generation uses a rollout policy and world model; action-free generation labels generated visual trajectories using inverse dynamics. Table 3 distinguishes wrist-view augmentation, action-conditioned Genie Envisioner/Ctrl-World, and action-free DreamGen/GigaWorld-0.
Go to primary source ↓e08PDF p. 5, Section 3.4, Equation (4), Table 4; p. 6, Section 3.4 continuation
Simulators use action-conditioned predicted outcomes and external reward. Categories include evaluation-only systems, sparse/dense-reward RL, and test-time adaptation. The text describes VLA-Reasoner using MCTS and AdaPower using test-time training.
Go to primary source ↓e09PDF p. 6, Section 4 and Table 5
Backbones are grouped as image/video generators, unified understanding/generation models, and representation models. Table 5 lists Stable Video Diffusion at 1.5B parameters, Emu3 at 8.5B, and V-JEPA 2 at 1B. These are backbone sizes, not complete system or training-compute measurements.
Go to primary source ↓e10PDF p. 6, Table 6, LIBERO header and SRPO (Online), SRPO (Offline), RynnVLA-002 rows
SRPO (Online) success rates (%) are Spatial 98.8, Object 100.0, Goal 99.4, Long 98.6, Avg. 99.2. SRPO (Offline) has Avg. 92.5 and Long 88.7. RynnVLA-002 has Spatial 99.0 and Avg. 97.4. The table supplies no seeds, trial counts, confidence intervals, or common training budget.
Go to primary source ↓e11PDF p. 6, Table 7, CALVIN ABC→D header and DreamVLA, UniVLA, TriVLA, GR-1 rows
DreamVLA has successive instruction-completion percentages 98.2, 94.6, 89.5, 83.4, 78.1 and Avg. Len. 4.44; UniVLA has Avg. Len. 4.41 and five-instruction success 75.1%; GR-1 has Avg. Len. 3.06. TriVLA is printed as 96.8, 92.4, 86.8, 83.2, 81.8 with Avg. Len. 4.37. These are the source values without correction.
Go to primary source ↓e12PDF pp. 6–7, Section 5; p. 7, Table 8 and caption
Metric families are video generation quality, flow accuracy, robot tasks, and benchmark metrics. The caption defines frequency stars, preferred-direction triangles, and GT as ground-truth reliance. PSNR/SSIM prefer higher values; MSE/LPIPS/FID/FVD prefer lower values. Success rate and average task progress concern policy outcomes; their GT cells are crosses.
Go to primary source ↓e13PDF pp. 7–8, Section 6; p. 8, Table 9 and caption
The inventory separates simulation from real-world datasets and records horizon, configuration, platform, camera, and dataset counts. It lists LIBERO with 6.5k trajectories and 20 scenes, CALVIN with 24k trajectories and four scenes, and Droid with 76k trajectories and 564 scenes. The caption defines BridgeData's superscript s as skills. Section 6 distinguishes temporal stability, environmental generalization, and realistic-dynamics validation for synthesizers/simulators.
Go to primary source ↓e14PDF p. 8, Section 6, Performance on Simulation Benchmarks
The authors characterize LIBERO and CALVIN ABC→D as near-saturated, citing SRPO (Online) at 99.2% and DreamVLA at 4.44 average task length, and argue these settings inadequately capture real-world embodied complexity.
Go to primary source ↓e15PDF p. 8, Section 7, all five future directions; Section 8
Directions are physical consistency, spatiotemporal (4D) perception, safety/reliability, long-horizon foresight, and failure-aware dynamics. Suggested approaches include physics priors, causal/counterfactual reasoning, geometric representations, uncertainty, memory, and negative/suboptimal experience.
Go to primary source ↓e16PDF p. 3, Contributions; pp. 4–8, Sections 3–8 and Tables 1–9; pp. 9–10, References
The survey provides conceptual definitions, literature taxonomies, backbone/benchmark inventories, and two cross-method performance tables. It introduces no experimentally trained model, original controlled ablation, or shared training/evaluation recipe for the compiled scores. No appendix follows the conclusion; references occupy PDF pp. 9–10.
Go to primary source ↓8.5 Primary sources
Towards Generalist Embodied AI: A Survey on World Models for VLA Agents ↗
PDF · 6,289 extracted words
Source fingerprint
e7e26f0d6e33443a532bd661c0119b434ff53fc63c17074d24b08bd4db51ba13