Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning
1. Paper overview
In one sentence: Training on ranked continuations of self-generated histories improves robot-video persistence, but increases sampling cost and leaves physical validity dependent on the reward. e02e03e04e10e13e15e19
| At a glance | What to know |
|---|---|
| Research problem | Source description Teacher-forced video prediction conditions on clean observations, whereas autoregressive deployment repeatedly feeds generated frames back into the model. Errors therefore change the next prediction’s input distribution. PersistWorld targets this exposure bias in action-conditioned robot simulation, where object disappearance and robot deformation can undermine policy evaluation. e02e19 |
| Core mechanism | |
| A key reported result | DROID autoregressive rollout fidelity: External: 0.86 / 24.42 / 0.070; wrist: 0.67 / 19.39 / 0.277. SSIM / PSNR (dB) / LPIPS. Ctrl-World validation split; recorded actions; 14 clips; averages over the rollout. Authors’ reproduced Ctrl-World: external 0.84 / 23.02 / 0.081; wrist 0.62 / 17.80 / 0.310. All six improve. Imported and reproduced baselines remain distinct. The approximately 11-second label conflicts with the 70-frame, 5-Hz description. e09 |
| Reading caution | Source description Candidate sampling increases training cost. Visual rewards enforce no explicit physics or geometry constraints, and the temporal plots still show deterioration at longer horizons. e19 |
Core contributions
- Source description
Compare independent futures from one frozen rollout history, while randomizing prefix depth to expose training to different accumulated-error regimes. e03
Figure 2. A shared imagined history makes candidate futures comparable. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the upper panel from the initial three-camera observation to the robot policy, then down the action arrow into the world model. Predictions return to the policy and history buffer. Below, S1 constructs a shared generated prefix, S2 forks K continuations, and S3 scores agreement with ground truth. S4 combines the green current predictor with the purple reference predictor. The purple arrows carry 1−β into the positive branch and 1+β into the negative branch. Equation 1 and Algorithm 1 (p. 38) distinguish these explicitly: line 35 defines the positive predictor; line 36 defines the negative predictor, subtracting β times the current prediction. Equation 2 and line 37 add both squared losses, weighted by r and 1−r. The blue minus marks predictor construction. e02e03e04e07e22e25
What it supports. The central change concerns the simulator’s training distribution. Candidate comparisons start from the same possibly corrupted history, so reward differences compare alternative continuations rather than unrelated scenes or rollout depths. The diagram also keeps the action-producing policy visibly separate from the observation-producing world model.
Where the evidence stops. Matching printed branch formulas does not establish the executable training procedure. Algorithm 1 omits the extreme-candidate filtering and KL term described elsewhere. Its EMA reference update also conflicts with Appendix A’s never-updated reference; gradient freezing within an update does not resolve this between-update discrepancy.
2. Motivation
2.1 The problem and the proposed response
Teacher-forced video prediction conditions on clean observations, whereas autoregressive deployment repeatedly feeds generated frames back into the model. Errors therefore change the next prediction’s input distribution. PersistWorld targets this exposure bias in action-conditioned robot simulation, where object disappearance and robot deformation can undermine policy evaluation. e02e19
2.2 What this reading follows
A robot simulator becomes unreliable when its imagined bowl disappears before the next action arrives. PersistWorld addresses that failure by changing how an existing video world model is post-trained. It first lets the model accumulate its own mistakes, then compares several possible continuations from the same history. Recorded multi-camera trajectories provide the reward, and contrastive denoising updates the simulator toward better continuations. The figures below trace that mechanism, examine whether improvements survive foreground masking, and separate competitive ablations from stronger prose claims. The final comparison asks a different question: whether a more faithful-looking simulator also estimates real policy performance accurately. e02e03e04e10e13e15e19
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 Benchmarks & simulators. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The foundational-work, neural-simulator and training-optimization placement fits. PersistWorld updates a visual simulator conditioned on externally supplied actions. It neither jointly predicts actions and video nor extracts actions through inverse dynamics. The recorded architecture/prediction quadrant “Not applicable” therefore fits this contribution; it is not an unverified One Model assignment. e02e03e15
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 drifting history into a fair comparison
Imagine two generated clips at different rollout depths. The later clip may receive a worse absolute score simply because its history is already damaged; comparing those scores would mix continuation quality with starting-state difficulty. PersistWorld instead generates one prefix and freezes its history before branching independent candidates. Each continuation starts from the same observation history and action context, with a private buffer thereafter. Varying prefix depth changes which corruption regimes the model encounters across updates, while group normalization makes each update depend on relative quality within that shared context. At deployment, the predictor still follows the ordinary autoregressive loop; the candidate group is a training device. Reader analysis: the fair-comparison rationale is stronger than Table 5’s evidence for any one prefix schedule being best. e02e03e05e13
5.2 Use rewards to change a denoising update
The key distinction is between the frames used to judge a sample and the latent used as its denoising target. Recorded frames determine visual reward. The sampled continuation supplies the clean latent that is re-noised for learning. Current and reference denoisers then form positive and negative mixtures around the reference. Equation 1 gives the reference coefficients as 1−β and 1+β, respectively; Algorithm 1 prints these in separate lines 35 and 36. Reward weight r emphasizes the positive mixture for a strong candidate; 1−r emphasizes the mirrored mixture for a weak candidate. Both weighted squared errors are added. This avoids differentiating through the sampling chain. Affine EDM conversion permits the clean-latent formulation. The theorem assumes unlimited capacity and an exact reference posterior mean; the evolving EMA reference and finite LoRA update should not be conflated with those assumptions. e04e07e20e25
5.3 Follow the evidence from images to executed behavior
The evaluation proceeds through increasingly task-specific questions. Full-frame metrics establish agreement with recorded videos; foreground masks ask whether object and robot regions also improve. Neither alone certifies useful control, especially because the training reward uses the same perceptual metric families. The human study supplies an additional judgment of visual realism, with the limits of eight in-house raters and selectable videos. Figure 8 then tests whether simulated progression tracks real policy differences, while retaining visible optimism. Finally, the spatial-picking experiment trains pi_0 using model rollouts and reports real-robot progression. That is actual execution evidence, but it is a small trial set with no reported uncertainty. Reader analysis: these results support a useful simulator improvement without establishing broadly accurate physical simulation. e05e09e10e11e15e18e19
5.4 Training and inference
During training
Use UNet LoRA rank 64 and alpha 64 plus a trainable action encoder; freeze other backbone and VAE parameters. The stated run uses 6,000 steps, Muon, learning rate 10^-4, batch size 64, and the top five plus bottom five rewards from each 16-candidate group. Training takes two days on eight NVIDIA H200 GPUs. e06
Appendix B marks KL regularization and an old-policy EMA rising to 0.5 by step 500 as defaults. Algorithm 1 updates that reference by EMA, although Appendix A describes a never-updated reference. It is frozen for differentiation within an update; the between-update descriptions conflict. e07
During inference
Use 50 Euler sampling steps without classifier-free guidance. Generated observations condition subsequent actions when evaluating a separate policy; prerecorded actions drive the principal fidelity benchmark. Candidate ranking and ground-truth rewards are post-training operations, not a reported deployment-time search controller. e02e08e09e15
5.5 Implementation flow
- Predict and feed back
Ctrl-World’s SVD-based model conditions on history and future pose targets, generates a multi-view clip, and encodes the prediction into the rolling history. A separate robot policy can supply actions; the world model generates observations. e02
- Share a corrupted prefix
Replicate the initial observed latent to initialize history. Sample P uniformly from 0 through 9 and generate P clips autoregressively. Freeze that history and branch K=16 independent continuations, each with its own buffer and F chunks. e03
- Score comparable futures
Average LPIPS, SSIM and PSNR over frames and equally across views. Compare with the recorded trajectory. Z-score rewards within the candidate group, clip advantages to [-1,1], and rescale to reward weights in [0,1]. e05
- Contrastive forward denoising
Add Gaussian noise to each generated clean latent and evaluate current and reference denoisers. Fit reward-weighted positive and negative mixtures to that generated latent. Ground truth supplies the reward; it is not this denoising target. No gradient through rollout sampling is required. e04e03
6. Experiments & results
PersistWorld post-trains Ctrl-World on its own imperfect video histories. It branches comparable futures, scores them against recorded multi-view trajectories, and applies reward-weighted contrastive denoising. The evidence supports better visual persistence and modest downstream policy benefits, with unresolved implementation details and no explicit physical-consistency guarantee.
6.1 Read the original evidence
Table 1. Post-training improves every reported camera metric against reproduced Ctrl-World. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the camera grouping: external cameras and wrist camera have different baselines and should be compared within their own blocks. SSIM and PSNR point upward; LPIPS points downward. The caption’s provenance symbols matter: stars identify values imported from the cited Ctrl-World paper, whereas the dagger identifies this paper’s reproduction. Compare each bold Ours row first with that daggered row. The external block also supplies WPE and IRASim comparisons; corresponding wrist values for those systems are absent. These are averages across the rollout, so the table does not show when an error first appears or whether any individual manipulation succeeds. e09
What it supports. Against reproduced Ctrl-World, external PSNR increases from 23.02 to 24.42 dB and wrist PSNR from 17.80 to 19.39 dB. LPIPS falls from 0.081 to 0.070 externally and 0.310 to 0.277 at the wrist. Together with the SSIM increases, this is consistent evidence of improved reference-video fidelity under the stated protocol.
Where the evidence stops. The source simultaneously says 14 clips, 70 frames, 5 Hz and approximately 11 seconds; the duration is unresolved. Its headline percentage improvements are not exactly recoverable from the rounded cells. No uncertainty across training seeds is supplied.
Table 2. Foreground gains weaken the explanation that only static backgrounds improved. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across each camera block twice: first through the three Object-Only columns, then through Robot-Only. The masks come from RoboEngine segmentation of interacting objects and the robot arm. The dagger again denotes the authors’ reproduced Ctrl-World baseline, as defined in Table 1. Follow metric directions rather than comparing the numerical sizes of different metrics. The paper averages these scores over the same 14-step rollout protocol. Compare methods within a given mask and camera; masked scores should not be subtracted from full-frame scores as though they measure identical pixel populations. This evaluation asks whether the model’s improvement remains visible in task-relevant image regions. e09e10e19
What it supports. Object LPIPS drops from 0.025 to 0.021 for external views and from 0.088 to 0.078 at the wrist. Robot-only PSNR increases from 17.62 to 19.25 dB externally and 25.50 to 27.24 dB at the wrist. Thus the measured benefit extends to both manipulated objects and the robot.
Where the evidence stops. Segmentation-conditioned image agreement is not a direct measure of force, contact, geometric feasibility or action error. The result supports foreground visual fidelity; stronger claims about physically correct interaction require additional measurements.
Figure 8. Better policy ordering coexists with optimistic simulated task progression. Original paper, p. 30 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each point using both legends: color identifies banana placement, block placement or marker rotation, and shape identifies GR00T N1.5, pi_0 or pi_0-FAST. The horizontal axis measures real progression; the vertical axis measures world-model progression. Values above the dashed diagonal indicate optimistic simulation. The solid blue line is a fitted trend, not the perfect-agreement target. Scores average five real and eleven simulated rollouts for each task-policy combination, awarding equal fractions for ordered substeps. Compare the summary correlation and MMRV across panels, then inspect how far points remain from the diagonal. Better ranking need not mean calibrated absolute performance. e15
What it supports. PersistWorld reports correlation 0.822 and MMRV 0.006, compared with Ctrl-World’s 0.796 and 0.053. That favors policy ordering in this small evaluation. Yet both panels place weak real-world performances substantially above perfect agreement, so the simulator still tends to make tasks appear easier than they are on the robot.
Where the evidence stops. Only three tasks and three policies are represented. The plotted p-values concern each correlation, not a test of the difference between correlations. These results do not by themselves establish a statistically reliable calibration improvement.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| DROID autoregressive rollout fidelity Ctrl-World validation split; recorded actions; 14 clips; averages over the rollout. | External: 0.86 / 24.42 / 0.070; wrist: 0.67 / 19.39 / 0.277. SSIM / PSNR (dB) / LPIPS | Authors’ reproduced Ctrl-World: external 0.84 / 23.02 / 0.081; wrist 0.62 / 17.80 / 0.310. All six improve. Imported and reproduced baselines remain distinct. The approximately 11-second label conflicts with the 70-frame, 5-Hz description. e09 |
| Foreground fidelity Same 14-step validation protocol; RoboEngine object and robot masks. | Object LPIPS external/wrist: 0.021/0.078; robot PSNR: 19.25/27.24. Object LPIPS; robot PSNR (dB) | Ctrl-World: 0.025/0.088 and 17.62/25.50. Gains survive background removal; image-mask fidelity does not verify contact dynamics. e10 |
| Human preference Blind randomized video pairs with ground truth; eight in-house experts who could choose videos. | 174 wins versus 43 losses; reported 80% preference. Preference votes | Original Ctrl-World. Supports perceived fidelity in this selected-rater study; repeated votes and self-selection limit population inference. e11 |
| Reward composition ablation DROID validation; 10-step rollouts; equal gradient-update counts. | Combined: 0.721 / 20.40 / 0.3387. Wrist SSIM / PSNR (dB) / LPIPS | LPIPS-only: 0.708 / 19.90 / 0.3242. Combining rewards trades some LPIPS for structural and pixel fidelity; it does not win every metric. e12 |
| Prefix sampling ablation DROID validation; 10-step rollouts; equal gradient-update counts. | Random 0–9: 20.40 / 0.3387. Wrist PSNR (dB) / LPIPS | Fixed 3: 20.51 / 0.3381; curriculum: 20.43 / 0.3412. Random sampling is competitive, but Table 5 contradicts the claim that it matches or exceeds every fixed baseline. e13 |
| Contrastive objective versus SF-style regression Validation rollouts of 10 interactions; regression baseline also uses generated histories. | 0.872 / 25.52 / 0.116. External SSIM / PSNR (dB) / LPIPS | SF-style: 0.868 / 25.35 / 0.119. An objective control, not full Self-Forcing reproduction; Figure 13 compares gradient steps, not wall-clock speed. e17 |
| Transfer to Bridge and RT-1 Separate SVD adaptation for 100,000 steps, then post-training; 10-interaction evaluation. | Bridge: 0.876 / 26.36 / 0.0957; RT-1: 0.865 / 26.63 / 0.1305. SSIM / PSNR (dB) / LPIPS | Adapted SVD: 0.864 / 25.20 / 0.1061 and 0.833 / 23.81 / 0.1683. Applicability after embodiment-specific training, not zero-shot transfer. e16 |
| World-model policy evaluation Three tasks × three policies; five real and eleven simulated rollouts per combination; ordered partial-progress scores. | 0.822 / 0.006. Pearson correlation / MMRV | Ctrl-World: 0.796 / 0.053. Ranking improves slightly, but both simulators overestimate real task progression. e15 |
| Real-robot spatial picking after policy finetuning Full-finetune pi_0 on 50 rollouts from each world model; four block positions; five real trials per block. | 0.667. Average task progression | Base policy: 0.350; Ctrl-World augmentation: 0.558. Small real-execution demonstration; this is progression, not binary success. Uncertainty is not supplied. e18 |
| Reference-based reward versus HPSv3 Reward substitution within PersistWorld on DROID; ours uses top-k=2. | 0.826 / 24.09 / 0.181; 41.3 s. SSIM / PSNR (dB) / LPIPS; seconds per training iteration | HPSv3: 0.802 / 22.58 / 0.193; 129.2 s. HPSv3 uses “Robot arm <task>”; this is not an end-to-end WorldCompass comparison. e23 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. The combined reward is a compromise across metrics. Original paper, p. 25 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each row changes the reward used for post-training, while the six columns evaluate the resulting model. Appendix B specifies DROID validation, 10-step rollouts and equal numbers of gradient updates. First inspect the metric aligned with each single-reward row: LPIPS-only has the smallest LPIPS, SSIM-only reaches the largest wrist SSIM, and PSNR-only gives the largest PSNR. Then inspect the other columns in the same row to see the cost of that specialization. The combined row uses negative LPIPS plus SSIM plus PSNR scaled by 1/32, with equal camera averaging. It should be read as a tradeoff rather than a sweep of column maxima. e05e06e12
What it supports. At the wrist, moving from LPIPS-only to the combined reward changes SSIM from 0.708 to 0.721 and PSNR from 19.90 to 20.40 dB, while LPIPS worsens from 0.3242 to 0.3387. This directly supports the practical argument for balancing complementary visual criteria, although it does not select a universally optimal balance.
Where the evidence stops. The source calls the combination a necessary regularizer. This table shows a metric tradeoff, not proof of necessity for every task. Equal update counts also do not establish equal rollout-generation cost, and no seed-level uncertainty is given.
Table 5. Random prefixes are competitive, but the table does not support uniform superiority. Original paper, p. 26 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The Strategy rows describe the prefix schedule used before candidate branching: a fixed prefix, a growing curriculum, or a fresh random draw from the indicated range. The main method selects Random Size 0–9. Compare that row against the fixed-size-3 row before accepting the accompanying prose. Fixed 3 has higher external PSNR and higher wrist SSIM and PSNR, plus lower wrist LPIPS; random 0–9 has lower external LPIPS and tied external SSIM. Against the curriculum, the random default improves four displayed metrics, ties wrist SSIM, and reduces wrist PSNR. All of these comparisons use the appendix’s 10-step validation protocol. e03e12e13
What it supports. The proposed schedule covers different levels of accumulated history corruption while remaining competitive on the reported averages. However, the displayed values leave room for a simpler fixed prefix to win several metrics. Reader analysis: this supports randomization as a workable default, not as an empirically dominant mechanism across every measured outcome.
Where the evidence stops. The paragraph below Table 5 says random 0–9 matches or exceeds every fixed baseline and beats the curriculum on five metrics. The displayed cells contradict those claims. No seed variability or per-depth breakdown resolves the discrepancy.
7. Analysis & limitations
7.1 What the evidence leaves open
Candidate sampling increases training cost. Visual rewards enforce no explicit physics or geometry constraints, and the temporal plots still show deterioration at longer horizons. e19
The 70-frame, 5-Hz and approximately 11-second descriptions are inconsistent. The claimed 14.0% external LPIPS and 9.1% wrist SSIM improvements also cannot be exactly reconstructed from Table 1’s rounded values; unrounded aggregates are absent. e09
Reward references are called “held-out ground truth” without separately identifying the post-training and evaluation splits; disjointness cannot be established. Tables omit seed variability, limiting interpretation of tiny ablation differences. e03e06e12e13
At beta=0, Equations 1–2 lose dependence on the learned predictor, contrary to the prose’s supervised-denoising interpretation of that limit. This is an algebraic reading, not an implemented correction. e04
7.2 Questions for discussion
- Does randomized prefix training improve late-horizon foreground fidelity once sample-generation cost and seed variability are controlled?
- Can improved policy ordering coexist with harmful simulator optimism when synthetic trajectories are used for training?
8. Reproducibility audit
8.1 Requirements and known gaps
Requirements include Ctrl-World weights/action interface, synchronized DROID trajectories, the validation split, perceptual metric implementations and RoboEngine masks. Preserve the 14-step main protocol separately from 10-step ablations. e02e06e09e10e12
Resolve beta, noise-level distribution, KL formulation/coefficient, LoRA target modules, metric preprocessing, seeds and precise split membership. Algorithm 1 omits the reported extreme-candidate filtering and KL term. e22
Table 7 uses H for prediction horizon, whereas Section 3.1 uses H for history and Section 3.3 uses F for branch chunks; their mapping is not explicit. Distinguish the evolving EMA reference from Appendix A’s fixed reference. e21e07
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: Test whether prefix diversity specifically protects late rollouts
Reader-proposed check, not performed: compare fixed prefix 3, random 0–4 and random 0–9 using the same checkpoint, disjoint training/evaluation trajectories, reward weights, group filtering and reference-update rule. Repeat independent seeds. Report both equal-gradient-update and equal-generated-frame budgets, because longer prefixes consume more sampling. Evaluate full-frame and object/robot-masked metrics by rollout depth under separate 10-step and 14-step protocols. The specific prediction is that wider prefix exposure helps the later, more corrupted histories. If random 0–9 has no reproducible late-depth advantage after budget control, Table 5’s competitive aggregate scores do not establish that depth diversity is responsible for the benefit. e03e06e07e09e10e12e13
Check 2: Measure policy calibration separately from policy ordering
Reader-proposed check, not performed: reproduce the three task progression rules and policy interfaces, then add held-out initial scenes and more real trials. Pair each real rollout condition with repeated simulations from Ctrl-World and PersistWorld, keeping actions and observations aligned. Measure Pearson correlation and MMRV alongside signed simulation-minus-real progression and absolute progression error, with uncertainty grouped by task and initial condition. A favorable ranking result should be judged separately from optimism: if correlation improves but signed overestimation persists, the simulator is better for ordering policies without being calibrated for absolute performance. Use the same distinction when testing whether synthetic-rollout finetuning improves real spatial-picking progression. e02e15e18
8.3 Reading coverage
Visual audit: Rendered and visually inspected the title/author page and all 28 declared pages supporting retained method, training, numerical, evaluation and reproduction claims. Cross-checked Figure 2 S4 against Equations 1–2, Equations 10–13 and Algorithm 1 lines 35–37: the positive reference coefficient is 1−β, the negative reference coefficient is 1+β, and both weighted squared errors are added. Page 38 was also inspected in a fresh 201-DPI render. The separate fixed-reference versus EMA discrepancy is disclosed. Inspected Figures 3–8 and 11–14, Tables 1–14, the relevant derivation pages and Algorithm 1. All six final crops were individually viewed; narrow ablations were rendered at 400 DPI, the other crops at 300 DPI, without upscaling or scientific alterations. Table 1 retains its short caption because it defines the imported/reproduced baseline symbols. Figures 1, 9 and 10 are outside the visual pass; their captions and the complete supplied text were read. No project videos, external code or separate supplements were inspected.
PDF pages inspected for this edition: 1, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 19, 20, 23, 24, 25, 26, 27, 28, 29, 30, 31, 34, 35, 36, 37, 38. 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 and abstract
- 1 Introduction
- 2 Related work
- 3 Improving Robot World Models with Reinforcement Learning
- 3.1 The Closed-Loop Gap in Autoregressive World Models
- 3.2 Online RL Post-Training via Reward-Contrasted Denoising
- 3.3 Adapting Group-Relative Training to Autoregressive Video
- 3.4 Visual Rewards for Multi-View Video Clips
- 4 Experiments
- 5 Conclusion, limitations and acknowledgements
- References
- Appendix A.1–A.2: EDM objective and theoretical analysis
- Appendix B: all ablation studies
- Appendix C: world-model policy evaluation
- Appendix D: human preference study
- Appendix E.1–E.3: inference, additional results and Algorithm 1
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.
- Identity: PDF p. 1 identifies arXiv:2603.25685v2 [cs.RO], dated 4 September 2026. Its exact title and all four authors match the catalog. The catalog submission date is 26 March 2026; this report reviews the supplied v2, not the original submission. No earlier version or revision comparison was supplied.
- All eight supplied text chunks, covering the complete 38-page paper including references and appendices, were read. Visual inspection covers the pages declared in the illustrated edition; Figures 1, 9 and 10 were read through extracted captions but their page images were not inspected.
- The supplied extraction does not reconstruct figure images; this limitation was addressed for the reported evidence by inspecting the retained PDF pages and six original crops.
- Separate supplemental material availability has not been fully verified. No separate supplement was supplied.
- Project videos and code were not inspected; no experiments were reproduced. The printed project link establishes provenance only, not verified code availability.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block and arXiv margin
Exact title, four authors, Czech Technical University affiliation, and arXiv:2603.25685v2 dated 4 September 2026 are visible.
Go to primary source ↓e02PDF p. 5, Figure 2 top and caption; p. 6, Section 3.1 Model overview
Separate robot-policy actions condition the world model; H=6 history latents, past poses and L=5 future seven-dimensional EEF targets produce three-view frames fed back into history.
Go to primary source ↓e03PDF pp. 8–9, Section 3.3 S1–S4; p. 38, Algorithm 1 lines 3–17, 30–40
A replicated initial latent seeds a random 0–9-clip prefix; 16 candidates use private histories and F continuation chunks; LoRA/action-encoder updates follow reward ranking.
Go to primary source ↓e04PDF p. 7, Equation 1; p. 8, Equation 2 and following paragraph; p. 20, Equations 10–13; p. 38, Algorithm 1 lines 31–37
Predictor branches mix reference/current outputs; reward-weighted errors target generated clean latents after forward noising. The prose also interprets the beta-to-zero limit.
Go to primary source ↓e05PDF p. 9, Section 3.4 Equation 3; p. 10, Equations 4–7; p. 11, reward weights
Metrics average over frames/views; negative LPIPS plus SSIM plus scaled PSNR forms the reward. Group z-scores are clipped and rescaled.
Go to primary source ↓e06PDF pp. 10–11, Section 4 Implementation details and Dataset
LoRA rank/alpha 64, action-encoder training, frozen other parameters, 6,000 Muon steps, learning rate 10^-4, batch 64, group 16 with top/bottom five, and eight H200 GPUs for two days are stated.
Go to primary source ↓e07PDF p. 19, Appendix A Notation; p. 28, Tables 9–11 and Old-policy EMA schedule; p. 38, Algorithm 1 line 40
Appendix A says the reference never updates; the algorithm updates it by EMA. Table 10’s default rises to 0.5 by step 500; Table 9 marks KL as used. Learning-policy EMA is separately ablated.
Go to primary source ↓e08PDF p. 34, Appendix E.1 Inference Details
Euler sampling uses 50 steps without classifier-free guidance.
Go to primary source ↓e09PDF p. 11, Section 4 rollout protocol and improvement percentages; p. 12, Table 1, all rows and caption
Table 1 contains camera-specific averages and distinguishes imported (*) from reproduced (dagger) baselines. The protocol says 14×5=70 frames, approximately 11 seconds, at 5 Hz.
Go to primary source ↓e10PDF p. 12, Object- and Robot-Centric Evaluation; p. 14, Table 2, all camera/mask columns
RoboEngine masks isolate objects and robot pixels. All listed masked metrics improve over reproduced Ctrl-World.
Go to primary source ↓e11PDF p. 12, Human Preference Study; p. 13, Figure 6; p. 31, Appendix D Study design and UI Design
The source reports 174 versus 43 votes, about 80% preference, eight in-house expert users, video self-selection, randomized placement and moderate inter-rater agreement.
Go to primary source ↓e12PDF p. 25, Appendix B protocol and Table 4; p. 26, Combined reward paragraph
Ablations use 10-step validation rollouts and equal gradient-update counts. Individual rewards win their corresponding metrics; combined reward trades among them.
Go to primary source ↓e13PDF p. 25, Appendix B protocol; p. 26, Table 5 and Prefix sampling strategy paragraph
Table 5 compares fixed 3/6, curriculum and random 0–9/0–4 prefixes. Random 0–9 is worse than fixed 3 in several columns despite the accompanying superiority claim.
Go to primary source ↓e14PDF pp. 26–28, Tables 6–11 and corresponding Appendix B paragraphs
The source ablates group size/filtering, horizon, camera weighting, KL and reference/learning-policy EMAs; alternate filtering improves metrics over the selected default.
Go to primary source ↓e15PDF p. 29, Figure 7, task definitions and Table 12; p. 30, protocol and Figure 8
Banana placement, block placement and marker rotation use pi_0, pi_0-FAST and GR00T N1.5. Five real and eleven model rollouts per combination yield ordered partial-progress averages; Figure 8 reports correlation/MMRV and optimism.
Go to primary source ↓e16PDF p. 35, Generalization beyond DROID; p. 36, Table 14
SVD is adapted for 100k steps to Bridge and RT-1 before post-training. Table 14 gives 10-interaction metrics; IRASim rows contain only imported PSNR, with other metrics absent.
Go to primary source ↓e17PDF p. 12, Comparison to Self-Forcing style baseline; p. 36, same comparison and Figure 13
The SF-style control regresses ground-truth latents from generated histories. Ten-interaction camera metrics and weighted reward versus gradient steps are reported.
Go to primary source ↓e18PDF pp. 36–37, Policy Improvement and Figure 14
The authors full-finetune pi_0 on 50 rollouts from each world model and test four block commands with five real trials per block; progression is 0.350/0.558/0.667.
Go to primary source ↓e19PDF p. 15, Limitations and Future Work; p. 14, Figure 5; p. 34, Figure 11
Candidate-generation overhead and absent explicit physical/geometrical constraints are acknowledged. Wrist/external temporal plots show both models deteriorating.
Go to primary source ↓e20PDF p. 20, Appendix A.1 Equations 8–13; p. 23, Theorem 1; p. 24, Algebraic steps 3–4
Affine EDM preconditioning supports branch construction. The ideal minimizer assumes unlimited data/capacity and an exact reference posterior mean.
Go to primary source ↓e21PDF p. 6, H and L definitions; p. 9, S2 definition of F; p. 27, Table 7 and Prediction horizon paragraph; p. 38, Algorithm 1 Require block
Main-text H denotes history, L generated frames, and F branch chunks. Table 7 instead labels its 1/3 horizon H as generated frames; the correspondence is unspecified.
Go to primary source ↓e22PDF pp. 10–11, Implementation details; pp. 27–28, KL discussion and Table 9; p. 38, Algorithm 1
Beta/noise distribution remain symbolic; concrete KL formulation/coefficient, full preprocessing/split/seed configuration and LoRA targets are omitted. Algorithm 1 loops over all candidates without the stated extreme filtering.
Go to primary source ↓e23PDF pp. 34–35, Ours vs. WorldCompass discussion; p. 35, Table 13
HPSv3 substitution uses “Robot arm <task>”. Table 13 lists ours with top-k=2 and 41.3 versus 129.2 training seconds per iteration alongside visual metrics.
Go to primary source ↓e24PDF p. 13, Figure 4 axes and caption
Paired metric histograms report about 98% improved samples; the LPIPS panel plots negative delta LPIPS, so positive values favor PersistWorld.
Go to primary source ↓e25PDF p. 5, Figure 2 S4 reference/current arrows; p. 7, Equation 1; p. 8, Equation 2; p. 20, Equations 10–13; p. 38, Algorithm 1 lines 35–37
Figure 2 assigns reference coefficients 1−β to the positive predictor and 1+β to the negative predictor. Equation 1 and Equations 10–11 use those same combinations. Algorithm 1 line 35 prints (1−β) times reference plus β times current; line 36 prints (1+β) times reference minus β times current. Line 37 adds the two squared errors with weights r and 1−r.
Go to primary source ↓8.5 Primary sources
Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning ↗
PDF · 12,883 extracted words
Source fingerprint
ae26eea650fde9422f6444728a97f5734618ec1089d9f9e38eb4188cfd3393e7