HaWMPO: Hallucination-Aware World Model-based Policy Optimization for Generalist Robot Policy
1. Paper overview
In one sentence: A learned hallucination score improves some imagined-rollout policy updates by discounting unreliable rewards, but the benefit depends on penalty strength and the score remains a proxy for physical validity. e-overviewe-ham-architecturee-reward-softe-liberoe-ham-validatione-robote-penaltye-ablation
| At a glance | What to know |
|---|---|
| Research problem | Source description Imagined rollouts reduce dependence on physical interaction, but accumulated visual prediction errors can become false rewards. The policy may exploit these errors. HaWMPO asks whether action-conditioned reliability estimates can improve the learning signal without modifying the simulator during policy optimization. e-overviewe-world-rollout |
| Core mechanism | Source description HAM learns continuous chunk reliability from a composite visual, geometric and motion proxy, using observed futures only to construct training targets. e-ham-targete-ham-architecture |
| A key reported result | LIBERO three-suite policy success: Spatial 77.2; Goal 61.6; Object 52.2; reported average 63.7. Success rate (%). Spatial, Goal and Object; one-shot SFT initialization and 200 RL steps; fixed world models, WoVR* without PACE. Simulation evaluation trial counts and uncertainty are unspecified. Average: base 48.7, WMPO 56.8, WoVR* 60.9. Differences are +15.0, +6.9 and +2.8 percentage points. WoVR* wins Goal, 64.0 versus 61.6. Best reported average, not uniform superiority. Retain Table 1 unchanged despite the penalty-selection conflict. e-traininge-liberoe-penaltye-ablation |
| Reading caution | Source description The authors identify incomplete hand-designed hallucination supervision, mostly short-horizon evaluation and limited adaptability from separately trained modules. Plausible generated frames do not establish accurate contact dynamics. e-limitationse-world-rollout |
Core contributions
- Source description
HAM learns continuous chunk reliability from a composite visual, geometric and motion proxy, using observed futures only to construct training targets. e-ham-targete-ham-architecture
- Source description
Reward-Soft incorporates hallucination scores into GRPO rewards. The paper evaluates both human-labeled hallucination detection and downstream policy success, which test different parts of this proposal. e-reward-softe-grpoe-ham-validatione-libero
Figure 2. HAM evaluates an imagined chunk in the context of the action that produced it. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the blue visual path and green action path from left to right. Predicted frames supply frame embeddings; the current and initial anchor images provide context tokens. Section 3.1 specifies frozen DINOv3 visual encoding and action-to-visual cross-attention inside the four fusion layers. The diagram then leads to a learnable-query head, feed-forward processing, an MLP and a sigmoid score. The current image tracks the local scene, while the anchor is the initial rollout observation. Neither is an observed future. The score direction is explicit in the caption and Equation (1): a higher hallucination score means lower predicted reliability. e-ham-targete-ham-architecturee-world-rollout
What it supports. The mechanism uses both visual evidence and commanded actions to assess generated transitions. During training, observed future sequences supply comparison targets; during imagined rollout, the assessor works from generated frames and conditioning inputs alone. This makes the score available before the imagined experience contributes to a policy update.
Where the evidence stops. The supervision combines four quality metrics and is explicitly a proxy for physical validity. A sigmoid output does not establish calibration as a probability of failure, and successful-looking imagery alone does not verify the commanded motion.
2. Motivation
2.1 The problem and the proposed response
Imagined rollouts reduce dependence on physical interaction, but accumulated visual prediction errors can become false rewards. The policy may exploit these errors. HaWMPO asks whether action-conditioned reliability estimates can improve the learning signal without modifying the simulator during policy optimization. e-overviewe-world-rollout
2.2 What this reading follows
A video world model can make a robot action look successful even when its predicted transition is wrong. HaWMPO places a learned reliability assessor between imagined experience and policy optimization. The assessor sees the proposed action and generated frames; its score changes the reward used by GRPO. Read the architecture first, then separate three questions: whether the assessor recognizes hallucination, whether its reward modulation improves simulated task success, and whether the resulting policy transfers to physical execution. The tables support promising but uneven gains. A conflict between the stated penalty-selection procedure and the reported Goal result deserves attention before reproduction. e-overviewe-ham-architecturee-reward-softe-liberoe-ham-validatione-robote-penaltye-ablation
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 | Not assigned |
| Architecture | Not assigned |
| Prediction paradigm | Not assigned |
| Quadrant | Not assigned |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Insufficient evidence to decide
The supplied taxonomy is entirely Not assigned, so no existing quadrant can be confirmed. Architectural evidence supports separate world-model-assisted policy optimization: a VLA predicts actions, a frozen video model predicts observations, and separate heads/models score reliability and reward. This is neither a demonstrated unified One Model nor inverse-dynamics action extraction; no catalog fields are changed. e-overviewe-grpoe-world-architecturee-world-rollout
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 1. Keep the learned simulator separate from the action policy
Start with the imagined environment described in Appendix A.5. The VLA consumes the current observation and instruction and proposes eight actions. A separate action-conditioned Wan2.2 model receives those actions, associated action context, a fixed reference image and four recent frames. Its VAE and diffusion Transformer generate future visual observations; the implementation decodes thirteen frames but returns only the eight future frames. The last four generated frames become the next visual context, while the reference stays fixed. Within a chunk, generation proceeds jointly through latent denoising; across chunks, predictions become inputs to later predictions. This feedback is what lets the policy gather synthetic experience, but also lets visual errors accumulate. The world model stays frozen during policy optimization and does not itself emit rewards or robot actions. e-world-architecturee-world-rolloute-grpo
5.2 2. Turn a quality proxy into a different learning signal
HAM is trained against one minus the mean of four normalized quality scores, so its target increases as a predicted chunk diverges from the recorded future. The model learns to estimate this quantity without access to that future at rollout time. Reward-Soft then applies Equation (3), R̂=(1−αH)R, before group-relative advantages are formed. Figure 1 abbreviates this as Multiplication; the actual factor is 1−αH, not H itself. My interpretation is that the central intervention changes how imagined experiences rank as learning examples, rather than repairing their pixels. This distinction matters: a correct score can reduce a misleading update while leaving the simulator's transition wrong. The paper's integrated ablation tests the resulting pipeline, but does not by itself isolate the benefit of learned scores from other reward transformations. e-overviewe-ham-targete-ham-architecturee-reward-softe-grpoe-ablation
5.3 3. Separate recognition, optimization and physical execution
The evidence forms three distinct tests. Table 3 asks whether HAM ranks human-labeled hallucinations; its bootstrap intervals accompany descriptive comparisons on fifty chunks. Tables 1 and 4 ask whether policies trained with its reward modulation achieve higher LIBERO success; the answer varies by suite and penalty. Table 2 then asks whether the resulting policy succeeds in actual G1 manipulation, with twenty trials per task and method. These tests complement one another, but none substitutes for the others. My reading is that the strongest defensible claim is improved reported average success under this particular fixed-simulator protocol, with preliminary physical transfer. Before extending that claim, resolve the α-selection inconsistency, establish detector data separation, and evaluate repeated runs. The appendix's plausible generated videos cannot independently verify contact dynamics or task execution. e-ham-validatione-liberoe-ablatione-penaltye-robote-world-rollout
5.4 Training and inference
During training
OpenVLA-OFT starts from one expert trajectory per task for SFT. Compared RL policies receive 200 optimization steps. World-model methods train their simulator once and freeze it; reproduced WoVR* omits PACE, so it is not the complete original WoVR pipeline. e-traininge-liberoe-baseline
The reported setup uses Adam, backbone learning rate 2.0×10⁻⁵, a stated value-network rate 3.0×10⁻³, weight decay 0.01, gradient norm limit 1.0 and eight NVIDIA H200 GPUs. The value network's role is not explained in the presented group-relative objective. e-traininge-grpo
During inference
During imagined rollouts, the last four generated frames refresh context while the reference stays fixed. Frames are 256×256; LIBERO actions have three translation, three rotation and one gripper component. Denoising jointly generates a chunk; recurrence across chunks is autoregressive. Instructions enter the VLA, not the video model. e-world-architecturee-world-rollout
HAM needs no observed future during rollout. The deployed object is the post-trained action policy; the supplied pipeline does not specify deployment-time HAM filtering or video-search planning. Physical robot trials and generated videos are separate evidence. e-ham-architecturee-grpoe-robote-world-rollout
5.5 Implementation flow
- Build an action-conditioned simulator
The inherited WoVR simulator uses Wan2.2-TI2V-5B: a video VAE and a 30-block DiT with hidden dimension 3072 and 24 heads. Actions condition cross-attention through projected tokens and block modulation through grouped-action embeddings added to timestep embeddings. e-baselinee-world-architecture
- Construct reliability supervision
Generate chunks from recorded successful and failed trajectories. Normalize DINOv3 similarity, depth consistency, optical-flow trajectory consistency and MUSIQ quality, then invert their equally weighted mean. Task failure is not itself hallucination. Appendix percentile groups describe the proxy distribution; training uses continuous targets. e-ham-targete-proxy-analysis
- Regress the hallucination score
Frozen DINOv3 features, projected context and projected actions enter four fusion layers with action-to-visual cross-attention, self-attention and feed-forward processing. A learnable query, MLP and sigmoid yield the score; HAM is trained with mean squared error. e-ham-architecture
- Roll out and update the policy
Eight parallel imagined environments generate chunks. A separate reward model assesses generated observations; HAM estimates reliability. Reward-Soft modifies rewards before group normalization. The GRPO objective uses clipped policy likelihood ratios and reference-policy KL regularization to update the VLA. e-reward-softe-grpo
6. Experiments & results
HaWMPO post-trains OpenVLA-OFT inside a frozen video world model. A separate hallucination detector discounts rewards assigned to unreliable imagined action chunks before GRPO updates the policy. Table 1 reports 63.7% average LIBERO success, but gains vary by suite and the penalty-selection description is inconsistent. Physical testing provides preliminary evidence on two G1 tasks.
6.1 Read the original evidence
Table 1. The best average includes a lower Goal score than the strongest baseline. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down each suite column before looking at Avg. HaWMPO leads Spatial and Object, but WoVR* is bolded in Goal. The average therefore summarizes heterogeneous changes. Section 4.1 supplies the one-trajectory-per-task SFT initialization; Section 4.2 states the 200-step RL budget and fixed-world-model protocol. The asterisk is scientifically relevant: WoVR* is the authors' reproduction with PACE omitted, not a direct copy of the original method's headline score. Read the last row literally as 77.2, 61.6 and 52.2. Do not replace the Goal cell with the separate tuned-penalty result in Figure 4. e-traininge-liberoe-penaltye-ablation
What it supports. HaWMPO's reported average is 63.7%, against 48.7% for the base policy, 56.8% for WMPO and 60.9% for WoVR*. The corresponding gains are 15.0, 6.9 and 2.8 percentage points. This establishes the best reported average under the paper's restricted comparison protocol.
Where the evidence stops. Section 4.4 claims per-suite penalty selection, but this table matches the α=0.3 values discussed in Section 4.6. The protocol conflict remains unresolved. Simulation trial counts, repeated-seed variation and confidence intervals are not reported.
Table 2. Small physical trial counts provide a separate transfer check. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the parenthesized counts rather than the percentages. Every method receives 20 trials for each task, so one additional success changes a task's rate by five percentage points. The Tissue-to-Box task requires the tissue to remain inside the white box after release; the Headphone column abbreviates Headphone-on-Stand and requires the released headphones to remain supported. Compare methods within the same task column. Appendix A.4 repeats the protocol and identifies Figure 9 as actual robot execution. This table concerns physical trials, whereas Appendix Figure 10 separately depicts world-model-generated frames. e-robote-world-rolloute-limitations
What it supports. HaWMPO succeeds in 17 tissue trials and 15 headphone trials, compared with 15 and 12 for the base policy, and 16 and 13 for WoVR. Its mean is 80.0%, versus 67.5% and 72.5%. These counts support observed improvement on both tested physical tasks.
Where the evidence stops. The differences involve only a few trials, with no reported statistical significance. The paper calls the evidence preliminary. Two tasks on one robot setup cannot establish broad real-world generalization or long-horizon robustness.
Table 3. Human labels test hallucination recognition separately from proxy-target regression. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Both metric columns reward larger values; the bracketed ranges are 95% confidence intervals. The evaluation comprises 50 chunks from 50 trajectories across ten LIBERO tasks, with 18 human-labeled hallucinations and 32 non-hallucinated examples. Pay particular attention to the dagger note: DINO, trajectory and depth comparisons require corresponding observed future frames. HAM and MUSIQ do not, which matters when a rollout exists only in imagination. The intervals come from 2,000 trajectory-level bootstrap resamples. The separate regression errors reported below the table concern agreement with the composite training target, not these human labels. e-ham-validatione-ham-targete-proxy-analysis
What it supports. HAM achieves AUROC 0.9375 and AP 0.8952, the highest reported point estimates. Its AUROC interval is [0.8648, 0.9911] and AP interval [0.7660, 0.9885]. MUSIQ-only scores much lower, supporting the narrower finding that image quality alone is a poor hallucination indicator on this evaluation set.
Where the evidence stops. The authors explicitly state that significance of the differences has not been established. Strong target-regression agreement cannot replace human evaluation, and this small test does not establish robustness to new tasks or policy-induced distribution shifts.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO three-suite policy success Spatial, Goal and Object; one-shot SFT initialization and 200 RL steps; fixed world models, WoVR* without PACE. Simulation evaluation trial counts and uncertainty are unspecified. | Spatial 77.2; Goal 61.6; Object 52.2; reported average 63.7. Success rate (%) | Average: base 48.7, WMPO 56.8, WoVR* 60.9. Differences are +15.0, +6.9 and +2.8 percentage points. WoVR* wins Goal, 64.0 versus 61.6. Best reported average, not uniform superiority. Retain Table 1 unchanged despite the penalty-selection conflict. e-traininge-liberoe-penaltye-ablation |
| G1 physical manipulation Tissue-to-Box and Headphone-on-Stand; 20 trials per method per task; success requires stable placement after release. | HaWMPO: tissue 85% (17/20), headphones 75% (15/20); average 80.0%. Success rate (%) and successes/trials | Base: 75% (15/20), 60% (12/20), average 67.5%. WoVR: 80% (16/20), 65% (13/20), average 72.5%. Observed average gains are 12.5 and 7.5 percentage points; two tasks and small counts establish only preliminary transfer evidence. e-robot |
| HAM hallucination detection 50 chunks from 50 LIBERO trajectories, ten tasks; human labels: 18 hallucinated, 32 non-hallucinated. Confidence intervals use 2,000 trajectory bootstrap resamples. | HAM AUROC 0.9375 [0.8648, 0.9911]; AP 0.8952 [0.7660, 0.9885], 95% intervals. AUROC and average precision (AP), higher is better | DINO-only: 0.9062 / 0.8299; Depth-only: 0.8941 / 0.8644; MUSIQ-only: 0.3056 / 0.3750. DINO/depth require observed futures; HAM does not. Highest point estimates, without established statistical significance. Separate proxy-target MAE 0.0174, RMSE 0.0250 and Spearman 0.9310 measure target agreement, not physical validity. e-ham-validation |
| Integrated HAM and Reward-Soft ablation LIBERO, 200 steps, unified α=0.3; removing HAM also disables Reward-Soft. | Full versus removed module: Spatial 77.2 versus 69.2; Object 52.2 versus 49.6; Goal 61.6 versus 64.0. Success rate (%) | Differences: +8.0, +2.6 and −2.4 percentage points. Figure 4 separately reports Goal 66.2 at α=1. The combined module helps two suites at this checkpoint. This does not isolate detector architecture from reward shaping. e-ablatione-penalty |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 4. The best hallucination penalty differs across task suites. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal axis is training steps and the vertical axis is success rate in percent. The shared legend assigns blue circles to α=1, orange squares to α=0.3 and green triangles to α=0.1. Compare colors within a panel rather than comparing visual slopes across panels with different vertical ranges. Equation (3) explains the intervention: increasing α makes the factor 1−αH smaller for the same positive hallucination score. At 200 steps, the orange curve leads Object and Spatial, while blue leads Goal. These are policy-success curves, not measurements of HAM accuracy or physical dynamics error. e-reward-softe-penaltye-liberoe-ablation
What it supports. The plotted final values favor α=0.3 on Object and Spatial, at 52.2% and 77.2%, and α=1 on Goal, at 66.2%. A stronger penalty therefore does not uniformly improve learning. The authors interpret this as a tradeoff between suppressing unreliable experience and retaining useful imperfect rollouts.
Where the evidence stops. There are no uncertainty bands or repeated-seed statistics. Section 4.4 prints Object as 52.0%, unlike the plotted 52.2%, and its claim of per-suite selection conflicts with Table 1. Keep the plotted values distinct from that table.
Table 4. Removing HAM also removes Reward-Soft, so this is a combined-module test. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each suite as its own three-checkpoint experiment. Steps are 80, 120 and 200; w/o HAM is the removed-module baseline, Ours is the full method, and Δ is full minus removed success rate in percentage points. Section 4.6 makes the intervention explicit: disabling HAM also disables its reward modulation. The comparison therefore tests their combined contribution. Spatial favors the full method at all three checkpoints. Object reverses direction at 120 steps, and Goal favors the removed module throughout. The text specifies a common α=0.3 for the full method in this ablation. e-ablatione-penaltye-reward-soft
What it supports. At 200 steps, the combined module improves Spatial by 8.0 points and Object by 2.6 points, but reduces Goal by 2.4 points. The table supports a task-dependent contribution from hallucination-aware reward modulation, rather than a claim that inserting HAM always improves policy success.
Where the evidence stops. This intervention cannot distinguish the value of learned reliability estimates from generic reward reweighting. The separate α=1 Goal result motivates tuning, but it should not be substituted into this fixed-penalty comparison.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify incomplete hand-designed hallucination supervision, mostly short-horizon evaluation and limited adaptability from separately trained modules. Plausible generated frames do not establish accurate contact dynamics. e-limitationse-world-rollout
Section 4.4 says Table 1 uses per-suite α, but Table 1's Goal 61.6 matches α=0.3, while Figure 4 gives 66.2 for α=1. Section 4.6 explicitly fixes α=0.3. Section 4.4 also says Object 52.0 where the figure and tables show 52.2. These remain unresolved reporting discrepancies. e-liberoe-penaltye-ablation
Proxy-defined percentile separation is partly built into label construction. Human-label evaluation is more independent, but small; the source does not establish detector train/test separation or downstream simulation uncertainty. e-proxy-analysise-ham-validatione-libero
7.2 Questions for discussion
- Does HAM detect policy-induced dynamics errors on unseen tasks, or chiefly reproduce the metric mixture used for supervision?
- Would the reported average improvement survive predeclared penalty selection, repeated seeds and the complete WoVR PACE baseline?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the initialized VLA, action-conditioned simulator, reward model, recorded action/video pairs and four metric implementations. The appendix names DepthAnything-v3 and NDTW, but exact normalization bounds, HAM training split, reward-model training, diffusion schedule and full RL configuration remain unspecified. e-ham-targete-proxy-analysise-world-architecturee-world-rolloute-grpo
Proposed checks: evaluate HAM on trajectory/task-disjoint human labels; then compare learned, shuffled and constant hallucination scores against unmodulated rewards with the same frozen simulator and optimization budget. Report both advantage changes and actual LIBERO success, with penalty selection separated from final evaluation. e-ham-validatione-reward-softe-grpoe-liberoe-penaltye-ablation
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: Does HAM generalize beyond its composite labels?
Reader-proposed check, not performed: construct trajectory-disjoint and task-held-out evaluation partitions, keeping normalization bounds fixed from training data. Obtain blinded human hallucination labels for generated chunks and compare HAM with MUSIQ plus the reference-based DINO, depth and trajectory metrics under the same information allowances as Table 3. Report AUROC, AP and trajectory-bootstrap intervals separately from proxy-target regression error. If proxy regression remains strong but human-label ranking collapses on held-out tasks, the interpretation that HAM supplies broadly useful rollout reliability would be weakened. e-ham-targete-proxy-analysise-ham-validation
Check 2: Does the score's alignment matter more than generic discounting?
Reader-proposed check, not performed: keep the one-shot policy initialization, frozen world model, reward model, eight parallel environments and 200-step budget fixed. Compare raw rewards with Reward-Soft using learned H, H shuffled among chunks within a group, and a constant H matching the group's mean. Repeat for α=0.1, 0.3 and 1 with matched seeds; choose penalties on a separate validation set. Measure group-advantage changes and actual LIBERO success. A consistent advantage for aligned HAM scores over distribution-matched shuffled scores would support the reliability mechanism; equal gains would favor a generic reweighting explanation. e-reward-softe-grpoe-traininge-liberoe-penaltye-ablation
8.3 Reading coverage
Visual audit: Visually inspected the title/byline and version stamp on p. 1; pipeline, equations, algorithm and training/evaluation details on pp. 2–9; limitations on pp. 9–10; all appendix method, data-analysis and physical-evaluation pages 13–16. All Figures 1–10 and Tables 1–4 were viewed in context. Six final original crops were individually inspected, including complete legends, table headers and the relevant Table 2/3 notes. Pages 11–12 contain references and were read in the complete text, without separate image inspection. Figure 1's Multiplication block was checked against Equation (3): the actual factor is 1−αH, not H. Section 4.5 refers to qualitative examples as Figure 8, but the paired prediction/ground-truth examples are Figure 5 on p. 8; Figure 8 on p. 14 is PCA. This pass covers supplied PDF images, not external videos, code or supplements.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 13, 14, 15, 16. 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, author block and abstract (p. 1): arXiv:2609.09941v1, 9 September 2026
- 1 Introduction (pp. 1–3)
- 2 Related Work, including 2.1–2.2 (p. 3)
- 3 Methodology, including 3.1–3.3 and Algorithm 1 (pp. 3–6)
- 4 Experiments, including 4.1–4.6 (pp. 6–9)
- 5 Conclusion and 6 Limitations (pp. 9–10)
- References (pp. 10–12)
- Appendix A.1 Data Collection and A.2 Training Data Analysis (pp. 13–14)
- Appendix A.3 Baseline (p. 14)
- Appendix A.4 Real-World Experiments (p. 15)
- Appendix A.5 World Model as Simulator (pp. 15–16)
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 inspecting original PDF pages and all ten figures and four tables. All five supplied text chunks were read completely; reference-only pages 11–12 were read as text, without separate page-image inspection.
- Only the supplied v1 was reviewed. Its title and four-author byline match the catalog; no alternate revision or edition was supplied or compared.
- Code, external links, datasets and separate supplemental media 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/byline, affiliation lines and arXiv margin stamp
Title and four authors match the catalog. The stamp identifies arXiv:2609.09941v1 [cs.RO], 9 Sep 2026. Affiliations are Joy Future Academy, JD; School of Artificial Intelligence, Jilin University.
Go to primary source ↓e-overviewPDF pp. 1–2, Introduction and Figure 1 with caption
The pipeline links a VLA, action-conditioned world model, HAM and reward model; unreliable imagined transitions can bias policy learning. Figure 1's Multiplication block is shorthand clarified by Equation (3).
Go to primary source ↓e-ham-targetPDF pp. 3–4, Section 3.1, Equation (1) and paragraph after Figure 2
HAM uses predicted eight-frame chunks, actions, current and anchor images. Four normalized quality metrics yield one minus their mean as hallucination target. Observed successful and failed trajectories supply targets; the composite is a proxy and failure alone is not hallucination.
Go to primary source ↓e-ham-architecturePDF p. 4, Figure 2, Section 3.1 and Equation (2)
Frozen DINOv3, context/action projection, four fusion layers, a learnable query and sigmoid head regress chunk hallucination by MSE. Observed future images are needed for supervision, not imagined-rollout scoring.
Go to primary source ↓e-reward-softPDF p. 4, Section 3.2, Equation (3)
Reward-Soft multiplies average raw chunk reward by 1−αH. H is average predicted chunk hallucination; α controls modulation. The prose mistakenly calls the formula Equation (1).
Go to primary source ↓e-grpoPDF pp. 5–6, Algorithm 1, Section 3.3, Equations (4)–(7)
Eight parallel imagined environments supply action chunks. Modulated rewards are normalized within groups before a clipped ratio and reference-policy KL objective updates the VLA. Algorithm 1 repeats the incorrect Equation (1) cross-reference.
Go to primary source ↓e-trainingPDF p. 6, Section 4.1
OpenVLA-OFT receives one expert trajectory per task for SFT. Adam settings are backbone 2.0×10⁻⁵, value network 3.0×10⁻³, weight decay 0.01, gradient norm 1.0; hardware is eight NVIDIA H200 GPUs.
Go to primary source ↓e-liberoPDF p. 6, Section 4.2 and Table 1, all method rows and suite/Avg columns
All RL policies are evaluated after 200 steps; world models are trained once then fixed. WoVR* excludes PACE. HaWMPO reports 77.2/61.6/52.2 and average 63.7; base average 48.7, WMPO 56.8, WoVR* 60.9. Trial counts and uncertainty are not given here.
Go to primary source ↓e-penaltyPDF p. 7, Figure 4 with caption and Section 4.4; PDF p. 9, Section 4.6 penalty discussion
Figure 4 at 200 steps shows α=0.3 Object 52.2, Goal 61.6, Spatial 77.2; α=1 Goal 66.2. Section 4.4 instead prints Object 52.0 and claims per-suite α for Table 1, whereas Section 4.6 fixes α=0.3. Figure 4 has no uncertainty bands.
Go to primary source ↓e-robotPDF p. 7, Section 4.3, Figure 3 and Table 2; PDF p. 15, Appendix A.4 and Figure 9
Each method has 20 G1 trials per task with stable placement after release as success. Tissue counts: base 15, WoVR 16, HaWMPO 17; headphone counts: 12, 13, 15. Means are 67.5%, 72.5%, 80.0%. Figure 9 is identified as actual physical execution.
Go to primary source ↓e-ham-validationPDF p. 8, Section 4.5, Table 3 including dagger note, and Figure 5; PDF p. 14, Figure 8
Human detection uses 50 chunks/trajectories over ten tasks, with 18 positives and 32 negatives; 2,000 trajectory bootstrap resamples give 95% intervals. HAM AUROC/AP are 0.9375/0.8952; DINO 0.9062/0.8299, trajectory 0.8872/0.8069, depth 0.8941/0.8644, MUSIQ 0.3056/0.3750. Reference requirements, significance boundary and separate proxy regression metrics are explicit. The Section 4.5 reference to Figure 8 for qualitative frame comparisons is inconsistent: those examples are Figure 5, while Figure 8 is PCA.
Go to primary source ↓e-ablationPDF p. 9, Section 4.6 and Table 4, all suite/step rows
Removing HAM simultaneously disables Reward-Soft. At α=0.3 and 200 steps, full/removed success is Spatial 77.2/69.2, Object 52.2/49.6, Goal 61.6/64.0. Object is also worse at 120 steps, 52.8 versus 54.2; Goal is lower at all reported checkpoints.
Go to primary source ↓e-limitationsPDF pp. 9–10, Section 6
The authors limit coverage of their handcrafted proxy, acknowledge mostly short-horizon manipulation evaluation and underexplored long-horizon real-world scaling, and describe separate HAM/world-model training as limiting adaptability.
Go to primary source ↓e-proxy-analysisPDF pp. 13–14, Appendix A.1–A.2 and Figures 6–8
Data processing names DINOv3, DepthAnything-v3, MUSIQ and optical-flow trajectory consistency via NDTW. Metrics are normalized and equally weighted. Top 10%, middle 40%, bottom 50% define descriptive quality groups; continuous composite supervision is used. Metric figures visualize distribution overlap and proxy-defined separation.
Go to primary source ↓e-baselinePDF p. 14, Appendix A.3
HaWMPO inherits WoVR's action-conditioned Wan2.2 world model and omits PACE. The appendix claims PACE is orthogonal; no experiment establishing additivity is presented.
Go to primary source ↓e-world-architecturePDF p. 15, Appendix A.5, Model architecture and Inputs and outputs
Wan2.2-TI2V-5B uses a VAE and DiT with 30 blocks, hidden dimension 3072 and 24 heads. Actions condition token cross-attention and timestep modulation. Five RGB context frames plus eight action steps are supplied; LIBERO actions have seven components and frames are 256×256.
Go to primary source ↓e-world-rolloutPDF p. 16, Appendix A.5, Equation (8), Closed-loop imagined rollouts and Figure 10 with concluding paragraph
The simulator decodes 13 frames, returns eight futures and fixes conditioning latents. Four recent frames update while the reference is retained. The VLA receives language; the frozen world model receives images/actions. Reward and reliability are estimated separately. Generated visual plausibility does not guarantee contact dynamics or physical success.
Go to primary source ↓8.5 Primary sources
HaWMPO: Hallucination-Aware World Model-based Policy Optimization for Generalist Robot Policy ↗
PDF · 6,899 extracted words
Source fingerprint
b50a6b325a460026f65acdc349fb71256d4dde0f2d2a28ecad6855d596e9e306