WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RL
1. Paper overview
In one sentence: WoVR makes imagined reinforcement learning more useful by stabilizing the simulator, shortening unreliable prefixes and refreshing its policy coverage, while leaving residual model error and a distinct physical-robot data protocol. e-probleme-overviewe-policy-resultse-robot-resultse-real-setting
| At a glance | What to know |
|---|---|
| Research problem | Source description A visually plausible grasp can receive imagined success despite failing in execution. Closed-loop video prediction compounds its own errors, while RL changes the action distribution and can exploit those errors. WoVR asks how to obtain useful policy updates from an imperfect simulator. e-problem |
| Core mechanism | Source description The framework combines simulator stabilization, Keyframe-Initialized Rollouts (KIR), and Policy-Aligned Co-Evolution (PACE), addressing model errors at different points in the learning loop. e-overview |
| A key reported result | LIBERO policy success, one-trajectory SFT: WoVR 69.5; Spatial/Object/Goal/Long: 84.2/80.8/77.4/35.8. Average task success rate (%). Four ten-task suites; 2,500 actual-environment trajectories per suite for each RL method. Base 40.5; online GRPO 44.6; WMPO 50.9. The reported gain over SFT is 29.0 percentage points. This is executed benchmark performance, not imagined success. e-policy-resultse-pace |
| Reading caution | Source description The authors lack a formal hallucination-propagation analysis or regret bound, and substantially longer-horizon or mobile manipulation remains underexplored. e-author-limits |
Core contributions
- Source description
The framework combines simulator stabilization, Keyframe-Initialized Rollouts (KIR), and Policy-Aligned Co-Evolution (PACE), addressing model errors at different points in the learning loop. e-overview
- Author claim
The authors claim that controlling hallucination makes learned video simulators practical for VLA reinforcement learning; the evidence is comparative performance and diagnostics rather than a reliability guarantee. e-overviewe-author-limits
Figure 3. Actions control the video simulator through normalization and cross-attention, while anchored memory carries visual context. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read from the encoder on the left to the decoder on the right. The reference frame supplies persistent scene context; the four memory frames supply recent motion. Follow the two dotted paths from Actions: one joins the timestep embedding and conditions normalization and scaling, while the other enters cross-attention. After decoding, the reference and memory reconstructions are marked Dismiss; generated future frames feed the dashed blue loop back into memory. These are simulator predictions under policy actions. The separate VLA policy in Figure 2 supplies those actions. Appendix B.1 describes rectified-flow training and noise on memory context, while keeping the initial anchor clean. e-architecturee-overviewe-flowe-kire-context-ablation
What it supports. The mechanism combines global scene anchoring with local temporal context and two action inputs. It explains how a video generator becomes an autoregressive simulator for policy training. Table 5 separately tests reference, memory and context noise; this drawing alone does not demonstrate the causal benefit of both action-conditioning channels.
Where the evidence stops. Figure 3 labels eight noisy future frames, but Appendix B.3, Equations (12)–(15), indexes nine first-chunk outputs. The supplied PDF does not resolve the discrepancy. The crop preserves the original labels, and no exact common chunk length is inferred.
2. Motivation
2.1 The problem and the proposed response
A visually plausible grasp can receive imagined success despite failing in execution. Closed-loop video prediction compounds its own errors, while RL changes the action distribution and can exploit those errors. WoVR asks how to obtain useful policy updates from an imperfect simulator. e-problem
2.2 What this reading follows
A video simulator can draw a successful grasp even when the same interaction fails in the environment. Training a policy against that image can reinforce the error. WoVR addresses this feedback loop with an action-conditioned video model, keyframe starts near decisive interactions, and low-frequency updates using data from the improving policy. The strongest reading separates three questions: whether generated videos stay coherent, whether benchmark policies improve, and whether those improvements survive physical execution. The figures and tables below connect these questions without equating them. In particular, the robot experiments replace PACE data collection with trajectories from multiple supervised-policy checkpoints. e-probleme-overviewe-policy-resultse-robot-resultse-real-setting
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 | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Dual-system and Other mechanisms classification fits the architecture: a separate action-conditioned video simulator supplies imagined experience to a VLA policy. Actions are policy outputs and simulator inputs, rather than jointly predicted future/action tokens or inverse-dynamics extraction. PACE co-evolution does not make this one unified model; the outside-quadrants placement is supported. e-overviewe-architecturee-pace
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 Turn a video predictor into a feedback environment
Start with the division of labor in Figure 2. The VLA policy maps the instruction and observation to an action; the video model predicts what that action will make the camera see; the reward model judges the predicted observation. Feeding that image back to the policy closes an imagined control loop. Appendix B.1 trains the video component by matching rectified-flow velocities between noise and future latents, conditioned on actions and visual context. The clean first frame supplies an enduring scene reference, while mildly corrupted memory prepares the model for imperfect generated inputs. Policy optimization is a separate GRPO update on the resulting trajectories. At physical deployment the policy acts on camera feedback, and platform-specific pose conversion produces joint commands. Generated pixels are training transitions, not actuator commands. e-overviewe-architecturee-flowe-kire-control
Figure 4. KIR removes a drift-prone imagined prefix before the critical interaction. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the rows at the decisive object interaction. The top row is ground truth. The red middle row starts at the episode beginning, accumulates errors through its prefix, and reaches an apparent success that the source identifies as false. The green lower row skips that prefix and starts near the critical state, where the model correctly depicts failure. Dismiss refers to omitted early rollout frames, not a test that rejects faulty predictions. Appendix B.3 implements the start with the original first frame plus four observations near a fixed task-level keyframe index; subsequent generation follows the ordinary autoregressive simulator. e-kire-pace-kir-ablation
What it supports. The example explains why preserving a real local context can improve outcome prediction even without changing the model architecture. The associated two-suite policy ablation reports 82.5% with KIR versus 79.7% without it while retaining PACE. This supports usefulness under that experiment, rather than measuring a universal reduction in hallucinations.
Where the evidence stops. This is one qualitative failure case. The paper does not enumerate task keyframe indices or the rollout mixing proportion. Predicting failure correctly here does not show that every remaining success reward is trustworthy.
5.2 Understand what shortening the prefix changes
Imagine a policy that usually approaches a difficult contact late in an episode. Starting every simulator rollout from the first frame forces prediction through all earlier motions before the useful correction can be learned. KIR supplies the initial anchor and local observations near that contact, shortening the imagined causal chain. Appendix B.3 uses a task-level index based on the base policy’s typical failure region, rather than manually selecting every trajectory. Equation (1) also normalizes each trajectory by its valid steps up to first success; the authors argue that short critical segments receive more weight per step. Reader interpretation: KIR changes both the states encountered during training and their gradient weighting. Its measured gain therefore does not by itself distinguish reduced prediction error from better training-state selection. e-kire-pace-kir-ablation
5.3 Read co-evolution as a data intervention
An improving policy can leave the behavior distribution on which its simulator was fitted. PACE addresses this by collecting a limited new batch with the evolved policy, refining the world model, and resuming imagined optimization. LIBERO fixes the actual-environment budget at 2,500 trajectories per suite, split into 1,500 base and 1,000 aligned rollouts for WoVR. Table 6’s additional-base-data control is important because it asks whether distribution coverage matters beyond sample count. Figure 8 supplies a diagnostic comparison of imagined and evaluated success around the simulator transition; it does not show disagreement disappearing uniformly. Reader interpretation must also respect Appendix F.4: the robot experiments obtain behavioral diversity from multiple SFT checkpoints and omit PACE data collection. Their success supports imagined post-training, but is not a physical ablation of policy-aligned co-evolution. e-pacee-pace-kir-ablatione-alignment-curvese-real-setting
5.4 Training and inference
During training
World-model training uses rectified flow and mild noise on non-reference context while retaining a clean initial anchor. The main reward model is a ResNet-based binary classifier trained with cross-entropy and thresholded at 0.5. e-flowe-reward
A Qwen3-VL alternative predicts 11 progress levels from four frames and task text using LoRA and an MLP head. It learns faster initially but reaches similar final performance with approximately three-times slower rollouts; sparse rewards are used in all main experiments. e-reward
Physical-robot world models instead train on rollouts from multiple SFT checkpoints, with no PACE data collection. The best physically evaluated SFT checkpoint initializes RL. This is a material protocol difference from the LIBERO study. e-real-setting
During inference
During imagined interaction, the policy and world model alternate, and the reward model supplies learning signals. Deployment directly runs the optimized policy on physical observations. Delta translation, Rotation 6D and gripper outputs are converted to absolute poses and IK joint commands; Piper accumulates deltas from its initial state to reduce feedback-noise sensitivity. e-overviewe-robot-resultse-control
5.5 Implementation flow
- Predict action-conditioned futures
A Wan2.2-TI2V-5B video backbone receives action embeddings through timestep-conditioned normalization and cross-attention in place of text embeddings. Its VAE encodes the persistent initial frame and recent memory; the DiT predicts future latents, which are decoded and fed back for the next chunk. e-architecturee-flow
- Start near decisive interactions
KIR initializes a portion of imagined rollouts with the initial anchor plus four frames near a task-critical state. In practice, a fixed index per task approximates where the base policy approaches failure. Subsequent chunks use the normal autoregressive loop; KIR changes context initialization, not architecture. e-kir
- Optimize valid imagined segments
A learned reward classifier scores generated observations. GRPO uses group-relative advantages and averages each trajectory over valid timesteps up to first success, masking later steps. The authors argue that this increases the per-step weight of short, decisive segments; it does not verify that the first predicted success is real. e-kire-reward
- Refresh the simulator distribution
PACE trains WM_Base from base-policy trajectories, improves the policy inside it, then fits WM_Evo using additional evolved-policy rollouts. LIBERO allocates 1,500 base plus 1,000 aligned trajectories per suite. Policy-gradient rollouts are imagined, but simulator refinement still requires actual-environment data. e-pace
6. Experiments & results
WoVR post-trains a VLA policy inside an action-conditioned video simulator. Anchored visual memory stabilizes predictions, keyframe initialization shortens imagined prefixes, and low-frequency simulator refinement addresses policy drift. LIBERO and physical execution improve, but physical experiments use a different data protocol without PACE. Simulator reliability remains empirical.
6.1 Read the original evidence
Table 2. WoVR improves executed LIBERO success under both reported SFT initializations. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two shaded blocks separately: the top starts from one-trajectory SFT, while the bottom starts from full-trajectory SFT. Columns report percent success for Spatial, Object, Goal and Long, with their average at the right. Within each block, compare WoVR against its own SFT base and then against the two RL alternatives. The red delta row is the change from the base in percentage points. Section 4.2 gives each RL method 2,500 actual-environment trajectories per suite. WoVR divides them into 1,500 base-policy and 1,000 evolved-policy rollouts for simulator training, whereas online GRPO uses the interaction budget directly. e-policy-resultse-pace
What it supports. In the weaker initialization, WoVR reaches 69.5% average versus 40.5% for SFT and 50.9% for WMPO. LIBERO-Long rises from 13.8% to 35.8%, so the gain is not confined to easier suites. The stronger initialization also improves, with the table reporting 96.0% versus an 88.1% base.
Where the evidence stops. The prose on page 8 reports 95.9% for the full-trajectory average, conflicting with the table’s 96.0%. Both are disclosed. Equal actual-environment data budgets do not establish equal total compute, and the table provides no confidence intervals.
Table 3. Physical execution improves, with substantial residual difficulty on Piper. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The left group covers Franka banana placement, bread insertion and drawer opening. The right group covers Piper towel folding, cube placement and tomato placement followed by closing the box lid. Each cell is a success percentage from 30 independent trials per task. Compare methods within a task before reading the platform averages; the platforms solve different task sets. The red deltas are changes from the base policy. Appendix F.4 adds a crucial qualification: these world models use trajectories from multiple SFT checkpoints, and the best evaluated checkpoint initializes RL. The robot data collection does not use the PACE procedure evaluated on LIBERO. e-robot-resultse-hardwaree-controle-real-setting
What it supports. WoVR reaches a reported 80.0% average on Franka and 28.9% on Piper, exceeding both SFT and WMPO. These are physical outcomes rather than generated-video scores. Gains coexist with low absolute Piper performance: towel folding reaches only 20.0%, while cube and tomato tasks each reach 33.3%.
Where the evidence stops. Thirty trials per task provide limited precision, and no confidence intervals are reported. Different tasks and platform-specific action conversion confound a hardware-only comparison. The physical results also cannot independently establish a benefit from PACE, which was not used for their data collection.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO policy success, one-trajectory SFT Four ten-task suites; 2,500 actual-environment trajectories per suite for each RL method. | WoVR 69.5; Spatial/Object/Goal/Long: 84.2/80.8/77.4/35.8. Average task success rate (%) | Base 40.5; online GRPO 44.6; WMPO 50.9. The reported gain over SFT is 29.0 percentage points. This is executed benchmark performance, not imagined success. e-policy-resultse-pace |
| LIBERO policy success, full-trajectory SFT Same four suites and interaction budget; stronger SFT initialization. | Table 2: 96.0; accompanying prose: 95.9. Average task success rate (%) | Base 88.1; online GRPO 89.7; WMPO 92.4. The report retains the table value and discloses the unresolved discrepancy; no uncertainty estimates accompany these values. e-policy-results |
| Action-conditioned video prediction at 512 frames LIBERO; 3,000 training and 200 held-out trajectories; common chunk-wise autoregressive protocol. | 0.091 / 34.252 / 68.011 / 0.154; 23.0 FPS. LPIPS / FID / FVD / FloLPIPS (lower); FPS (higher) | OpenSora: 0.105 / 38.478 / 89.391 / 0.156; 7.00 FPS. WoVR uses five diffusion steps. Perceptual metrics and throughput do not directly certify contact physics or reward correctness. e-wm-resultse-metrics |
| Physical manipulation with OpenVLA-OFT Three tasks per robot; 30 independent trials/task; multi-task policies; checkpoint-diverse world-model data without PACE collection. | Franka 80.0; AgileX Piper 28.9. Platform-average success rate (%) | Base: 51.1/15.5; WMPO: 64.5/18.9. Physical gains are 28.9 and 13.4 percentage points. Piper remains difficult; different platform tasks prevent a controlled hardware comparison. e-robot-resultse-real-settinge-hardware |
| Physical manipulation with pi_0.5 Piper; Fold Towel, Pick Cube and Pick Tomato; Flow-SDE optimization; 30 trials/task. | 56.7. Average task success rate (%) | Base 34.4; reported gain 22.3 percentage points. Supports usefulness with a second policy backbone under the reported three-task setting. e-pi-results |
| Context-mechanism ablation LIBERO-Spatial; 1,500 training and 24 held-out trajectories; 512-frame rollout. | Full WoVR 73.493. FVD (lower) | No reference 123.502; one memory frame 86.042; no noisy context 77.284. Anchoring has a strong measured effect here. Noise is not uniformly superior: at 256 frames, no-noise FVD is 61.660 versus 63.948. e-context-ablation |
| PACE and KIR ablation One-trajectory SFT; Spatial and Object only; imagined policy optimization. | Full WoVR 82.5. Two-suite average success rate (%) | 1,500 base trajectories: 75.8; 2,500 base: 77.5; PACE without KIR: 79.7. The matched data-count comparison supports policy alignment; removing KIR costs 2.8 percentage points. This average is not the four-suite headline metric. e-pace-kir-ablation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 5. The fixed reference frame has the clearest long-horizon effect in this ablation. Original paper, p. 19 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each method occupies three rows, one for each rollout horizon. Compare the same horizon across variants rather than comparing adjacent rows within a variant. All four metrics have downward arrows: LPIPS describes frame perceptual discrepancy, FID and FVD compare feature distributions, and FloLPIPS incorporates motion discrepancy. Appendix D.1 uses only LIBERO-Spatial, with 1,500 training trajectories and 24 held-out trajectories. Begin with the 512-frame rows to inspect the longest reported feedback chain, then check whether the ordering survives at shorter horizons. The no-reference row removes the persistent anchor; the other variants reduce memory or disable noisy context during training. e-context-ablatione-metricse-architecture
What it supports. At 512 frames, full WoVR has FVD 73.493 versus 123.502 without the reference frame, 86.042 with one memory frame, and 77.284 without context noise. These values support anchored temporal context for this generation setting. They complement Figure 7’s selected examples of drift and disappearing objects.
Where the evidence stops. The improvement is not uniform: at 256 frames, no-noisy-context FVD is 61.660, lower than full WoVR’s 63.948. This small held-out, single-suite study measures video quality; it neither isolates dual-channel action injection nor directly quantifies executed policy success.
Table 6. Policy-aligned simulator data and keyframe starts contribute beyond simply adding base-policy rollouts. Original paper, p. 20 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Avg as an average over Spatial and Object only. The first two WoVR ablations differ in how many base-policy trajectories train the simulator: 1,500 versus 2,500. The full method also uses 2,500 total trajectories, but replaces 1,000 base-policy trajectories with data from the evolved policy. This provides a same-count comparison for PACE. The no-KIR row retains that PACE protocol while starting imagined rollouts at the episode beginning. All variants optimize the policy through imagined interaction. Therefore, the table separates a data-distribution change from an initialization change more directly than the headline comparison against a different video backbone. e-pace-kir-ablatione-pacee-policy-results
What it supports. Adding base-policy data raises the two-suite average from 75.8% to 77.5%. Full WoVR reaches 82.5%, five percentage points above the same-count base-data variant. Removing KIR lowers the average to 79.7%. Together these comparisons support both aligned simulator refinement and shorter task-critical imagined prefixes under the reported setup.
Where the evidence stops. Only two suites are reported, so 82.5% must not replace the four-suite headline average. The design controls trajectory count for the key PACE comparison, but does not document an exhaustive matched-compute, multi-seed factorial study or prove that hallucination has been eliminated.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors lack a formal hallucination-propagation analysis or regret bound, and substantially longer-horizon or mobile manipulation remains underexplored. e-author-limits
Success classification can still reward physically false predictions. Selected failure illustrations are qualitative; neither attractive videos nor higher success alone measures the frequency of simulator exploitation. e-probleme-rewarde-failures
Figure 3 labels eight noisy future frames, whereas Appendix B.3 indexes nine first-chunk outputs. The PDF does not resolve this indexing discrepancy. Reported success tables lack confidence intervals; physical results use only 30 trials per task. e-architecturee-kire-policy-resultse-robot-results
7.2 Questions for discussion
- Does KIR help because it shortens prediction depth, because it selects informative states, or both?
- Would physically collected evolved-policy data improve on the checkpoint-diversity protocol used for the robot experiments?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the policy/video backbones, task-matched rollout datasets, success labels and robot action conversion. Table 7 reports Franka/Piper world-model SFT steps of 750k/1.25M, episode limits of 160/224, and RL epochs of 50/60. Physical data use D435 RGB recording at 30/10 Hz respectively. e-architecturee-rewarde-real-settinge-hardwaree-control
Appendix A explains shared GPU scheduling, but does not identify hardware or a complete throughput configuration. Exact context-noise strength, task keyframe indices, KIR mixture, and a complete GRPO hyperparameter/seed configuration are not provided in the supplied implementation sections. These remain prerequisites for a faithful repeat. e-systemse-flowe-kir
Proposed checks: compare critical versus random keyframes at matched remaining horizons; independently compare evolved-policy versus extra base-policy data at fixed sample and training budgets. Measure actual success and imagined-versus-actual disagreement, not only generated-image quality. e-kire-pace-kir-ablatione-metrics
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: Separate keyframe relevance from shorter prediction depth
Reader-proposed, not performed: freeze the same world model, policy and reward classifier on held-out LIBERO-Spatial episodes. Compare contexts near the task-critical index with non-critical contexts drawn from matched time indices in other episodes of the same task; keep the remaining imagined horizon and sample count fixed. Include initial-state rollouts as a separate long-prefix reference. From matched simulator states, compare imagined success labels with actual-environment outcomes under the policy. Then repeat policy updates with equal generated-frame budgets and evaluate from ordinary episode starts. If critical and time-matched random starts perform equally, the special role of task-critical selection is weakened even if both beat long-prefix imagination. e-kire-context-ablatione-pace-kir-ablation
Check 2: Test alignment against an equal-budget data refresh
Reader-proposed, not performed: start both arms from the same world model trained on 1,500 base-policy trajectories and the same first-stage policy checkpoint. Refresh one simulator with 1,000 evolved-policy trajectories and the other with 1,000 additional base-policy trajectories. Match task balance, optimizer steps, reward model, KIR settings and second-stage imagined-rollout budget; repeat with independent seeds. Evaluate ordinary-start success and false-success disagreement on held-out evolved-policy trajectories before and after refresh. PACE’s alignment explanation predicts a larger reduction in disagreement and better actual-environment success for the evolved-data arm. Improvement only in imagined reward, or equal gains from base data, would weaken that explanation. e-pacee-pace-kir-ablatione-alignment-curves
8.3 Reading coverage
Visual audit: Visually inspected the title/version/author page, all eleven figures and all seven tables, and every retained method, training, evaluation and reproduction evidence page. Figure 3 action paths and feedback direction were cross-checked with Section 3.1 and Appendix B.1; Figure 4 prefix omission and first-success masking were checked against Section 3.2 and Appendix B.3. The eight-versus-nine frame discrepancy and the Table 2 versus prose average discrepancy are disclosed. All six final original crops were viewed for legibility and complete labels. References and appendix contents on pages 10–14 were read as text; external videos, code and separate supplements were not inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25. Appendix coverage: reviewed.
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
- Title, abstract and Sections 1–2 (Introduction; Related Work)
- Sections 3.1–3.3 (world model, KIR/GRPO, PACE)
- Sections 4.1–4.3 (world-model, LIBERO and physical-robot experiments)
- Sections 5–6 (Conclusion; Limitations)
- References and appendix contents
- Appendix A (GPU allocation)
- Appendices B.1–B.3 (world-model, reward and KIR implementation)
- Appendix C (evaluation metrics)
- Appendices D.1–D.3 (context, PACE/KIR and reward ablations)
- Appendix E (qualitative failure analysis)
- Appendices F.1–F.4 (hardware, tasks, action execution and training settings)
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 limitation was addressed by visually inspecting PDF pages 1–9 and 15–25, including all figures and tables; references and appendix contents on pages 10–14 were read in text.
- Identity/version scope: the title and all 14 authors match the catalog. The inspected title page identifies arXiv:2602.13977v2, 27 June 2026; the catalog submission date is 15 February 2026. No earlier edition was supplied, so revision contents were not compared.
- The title-page project link is wovr-rlinf.github.io, whereas the catalog projectUrl is wovr-corl.github.io; neither was accessed.
- Code, linked models, project videos and external supplements were not inspected. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title block and arXiv margin stamp
The exact title and all 14 catalog authors match. The artifact identifies arXiv:2602.13977v2 [cs.RO], 27 Jun 2026, and lists five affiliations.
Go to primary source ↓e-problemPDF p. 2, Introduction and Figure 1
Hallucination means disagreement between imagined and actual closed-loop outcomes. Autoregressive feedback and changing policy distributions can produce spurious success.
Go to primary source ↓e-overviewPDF p. 3, Figure 2 and caption; pp. 4–6, Sections 3.1–3.3
The architecture separates VLA policy, action-conditioned world model and reward model. KIR reshapes imagined interaction; PACE refines the simulator using evolved-policy trajectories.
Go to primary source ↓e-architecturePDF p. 4, Section 3.1 and Figure 3
Wan2.2-TI2V-5B receives actions through timestep-conditioned normalization and cross-attention replacing text embeddings. The context uses an initial anchor and memory. Figure 3 depicts one reference frame, four memory frames and eight noisy frames; decoded reference and memory frames are marked Dismiss.
Go to primary source ↓e-flowPDF pp. 15–16, Appendix B.1, Equations (2)–(5)
Future latents and Gaussian noise define a rectified-flow velocity target. Conditioning includes anchored visual context and actions. Non-reference context receives mild noise; its interpolation coefficient is only specified as close to one.
Go to primary source ↓e-kirPDF p. 5, Section 3.2, Figure 4 and Equation (1); p. 17, Appendix B.3, Equations (11)–(15)
KIR replaces repeated initial-frame context with the first frame and four observations near a fixed task-level pre-failure index. GRPO normalizes by valid steps up to first success. Appendix equations index nine first-chunk outputs; exact task keyframe indices, KIR mixture and complete GRPO configuration are not supplied.
Go to primary source ↓e-pacePDF pp. 5–6, Section 3.3; p. 7, Section 4.2 setup
PACE improves the policy in WM_Base, then refines the simulator using evolved-policy data. LIBERO uses 1,500 base plus 1,000 evolved-policy trajectories per suite; each comparison RL method receives a 2,500 actual-environment-trajectory budget.
Go to primary source ↓e-rewardPDF p. 5, reward paragraph; p. 16, Appendix B.2, Equations (6)–(10); pp. 21–22, Appendix D.3 and Figure 9
Main experiments use a ResNet-based binary classifier with a 0.5 threshold and binary cross-entropy. Qwen3-VL uses four frames and task text to predict 11 progress levels with LoRA and an MLP head. Dense rewards improve early learning but reach similar final performance with approximately three-times slower rollout.
Go to primary source ↓e-wm-resultsPDF p. 6, Section 4.1 setup; p. 7, Table 1, WoVR/OpenSora 512-frame rows and following paragraph
The evaluation uses 3,000 training and 200 held-out trajectories of 512 frames with common chunk-wise generation. WoVR reports LPIPS/FID/FVD/FloLPIPS 0.091/34.252/68.011/0.154 and 23.0 FPS; OpenSora reports 0.105/38.478/89.391/0.156 and 7.00 FPS. WoVR uses five diffusion steps.
Go to primary source ↓e-policy-resultsPDF p. 7, Table 2 and Section 4.2 setup; p. 8, Section 4.2 continuation
Each LIBERO suite has ten tasks. One-trajectory base/GRPO/WMPO/WoVR averages are 40.5/44.6/50.9/69.5 percent; full-trajectory averages are 88.1/89.7/92.4/96.0. The prose instead says 95.9. WoVR one-trajectory suite values are 84.2/80.8/77.4/35.8; full-trajectory values are 98.8/98.8/94.8/91.4. No confidence intervals are tabulated.
Go to primary source ↓e-robot-resultsPDF p. 8, Table 3 and Section 4.3 setup
Physical evaluation uses 30 independent trials per task and one multi-task policy per platform. Franka base/WMPO/WoVR averages are 51.1/64.5/80.0 percent; Piper averages are 15.5/18.9/28.9. WoVR Franka banana/bread/drawer values are 86.7/90.0/63.3; Piper towel/cube/tomato values are 20.0/33.3/33.3.
Go to primary source ↓e-pi-resultsPDF p. 8, final paragraph; p. 9, Table 4
Piper experiments with pi_0.5 and Flow-SDE use 30 trials per task. Mean success rises from 34.4 to 56.7 percent; towel/cube/tomato rates rise from 20.0/60.0/23.3 to 30.0/86.7/53.3.
Go to primary source ↓e-author-limitsPDF p. 9, Section 6
No formal hallucination-propagation characterization or regret bound is provided; much longer horizons and mobile manipulation remain underexplored.
Go to primary source ↓e-systemsPDF p. 15, Appendix A and Figure 6
Generation, neural simulation and policy training share GPU resources. Policy and simulator weights transfer at rollout-phase boundaries. The appendix illustrates scheduling without specifying GPU model or a reproducible throughput configuration.
Go to primary source ↓e-metricsPDF pp. 17–18, Appendix C, Equations (16)–(20)
LPIPS compares perceptual features; FID/FVD compare image/video distributions; FloLPIPS weights perceptual discrepancies by optical-flow differences; FPS counts generated frames per second.
Go to primary source ↓e-context-ablationPDF p. 18, Appendix D.1 setup; p. 19, Table 5; p. 20, Figure 7
The Spatial-only ablation uses 1,500 training and 24 held-out trajectories. At 512 frames full/no-reference/single-memory/no-noisy-context FVD is 73.493/123.502/86.042/77.284. At 256 frames no-noisy-context FVD is 61.660 versus full 63.948. Figure 7 illustrates drift and object disappearance.
Go to primary source ↓e-pace-kir-ablationPDF pp. 19–20, Appendix D.2 and Table 6; p. 21, KIR result paragraph
One-trajectory SFT averages on Spatial and Object are 75.8 percent with 1,500 base trajectories, 77.5 with 2,500 base, 79.7 with PACE but no KIR, and 82.5 for full WoVR. These are two-suite averages.
Go to primary source ↓e-alignment-curvesPDF p. 20, final paragraph; p. 21, Figure 8 and caption
Blue curves denote imagined training success, red actual-environment evaluation, and the vertical band the PACE transition. The authors interpret the change as simulator recalibration; disagreement is not uniformly eliminated across suites.
Go to primary source ↓e-failuresPDF p. 22, Appendix E and Figure 10
Examples describe unstable grasping, release and placement errors, and incomplete drawer, box-closing or towel interactions, with improvements after post-training. No failure-category counts are provided.
Go to primary source ↓e-hardwarePDF p. 23, Appendices F.1–F.2 and Figure 11
Six-DoF Piper uses head-view D435 RGB at 10 Hz; seven-DoF Franka uses fixed third-person D435 RGB at 30 Hz. Both have two-finger grippers. Tasks are banana-to-plate, bread-to-toaster, drawer opening, cube-to-box, tomato-to-box with lid closure, and towel folding.
Go to primary source ↓e-controlPDF pp. 24–25, Appendix F.3, Equations (21)–(26) and platform-specific updates
Delta EEF actions contain translation, Rotation 6D and gripper components. They are converted to absolute EEF poses and IK joint commands. Franka uses current-state delta conversion; Piper accumulates deltas from its initial state to reduce sensitivity to noisy feedback.
Go to primary source ↓e-real-settingPDF p. 25, Table 7 and Appendix F.4
Robot experiments omit PACE data collection, using multiple SFT checkpoints for world-model rollouts and the best physically evaluated SFT checkpoint for RL initialization. Franka uses 25 demos and 120 rollouts per task; Piper cube/tomato/towel uses 75/50/50 demos and 180/180/210 rollouts. Franka/Piper WM SFT steps are 750k/1.25M, episode limits 160/224, RL epochs 50/60.
Go to primary source ↓8.5 Primary sources
WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RL ↗
PDF · 11,505 extracted words
Source fingerprint
46eb8900c397d96282333f4110f6a30072fd07579eea8144c6990cadd50e1de9