Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling
1. Paper overview
In one sentence: GPC improves a frozen diffusion policy by ranking or refining its actions through a separate predictive world model, trading better manipulation outcomes for substantial inference cost. e01e02e03e06e12e16e18
| At a glance | What to know |
|---|---|
| Research problem | Source description Behavior cloning can compound errors after drifting away from demonstrated behavior. GPC asks whether learned predictive foresight can correct deployment decisions without retraining the policy. It assumes an action-conditioned dynamics model and an outcome reward or candidate-selection rule are available. e02e06 |
| Core mechanism | |
| A key reported result | Vision-based Push-T planning–latency tradeoff: RANK (K=100,M=0): 0.739 / 11.745; OPT (K=1,M=25): 0.791 / 39.061; RANK+OPT (K=10,M=25): 0.882 / 374.102. Mean IoU / seconds per decision. Table III; simulation IoU over 100 evaluation seeds; wall-clock seconds per decision. Behavior cloning: 0.642 / 0.457; RANK (K=50,M=0): 0.698 / 5.835. Combined planning gains 0.240 absolute IoU over BC at substantial cost. The table labels BC as K=1,M=1, conflicting with the no-refinement baseline elsewhere. e12e03e09 |
| Reading caution |
Core contributions
Figure 1. Use the policy to propose actions, then use predicted consequences to choose or improve them. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the blue generative-policy block and follow its output to the orange predictive-world-model block. In the upper panel, several proposed action sequences produce separate imagined trajectories; the star marks the highest reward, which here is the least negative value. Read horizontally for physical time. In the lower panel, successive rows represent optimization of one proposal, while horizontal position still represents time within a predicted rollout. The dashed gradient arrows return toward the proposal through the world model. Figure 1's arrows and stars agree with Algorithm 1 and Equations (2–5): the optimized variables are actions, and the policy remains frozen. e03e06e11
What it supports. The architecture gives GPC two ways to use extra computation. Ranking searches among policy samples, whereas refinement can move away from a sampled action chunk. Combining them supplies multiple starting points for optimization. The world model contributes foresight at inference rather than an auxiliary policy-training loss.
Where the evidence stops. The pictured futures are model predictions, not executed robot trajectories. A high predicted reward can be wrong. Gradient refinement requires a differentiable reward and world model; the paper's VLM selector supports ranking without supplying these gradients.
2. Motivation
2.1 The problem and the proposed response
Behavior cloning can compound errors after drifting away from demonstrated behavior. GPC asks whether learned predictive foresight can correct deployment decisions without retraining the policy. It assumes an action-conditioned dynamics model and an outcome reward or candidate-selection rule are available. e02e06
2.2 What this reading follows
A diffusion policy can propose several plausible robot motions without knowing which will work best from the current situation. GPC adds a second model that imagines the consequences of each proposal. A reward predictor then selects a candidate or supplies gradients that refine it before execution. This reading follows the separation between proposal generation, future prediction, and physical control, then checks the experiments against their data and runtime budgets. In the visual Push-T ablation, combining ranking and refinement gives the highest listed IoU, while the real-world evidence is limited to small trial sets. The supplied source is the March 2026 arXiv v4 accepted preprint. e01e02e03e06e12e16e18
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
Separate policy and predictive world-model networks support Dual-system. The world model conditions on proposed actions; it neither jointly emits future observations and executable actions nor infers actions from a desired future. Other mechanisms and Outside quadrants fit this inference-time planning design. The post-training category describes when enhancement occurs, not policy fine-tuning or demonstrated online RL. e03e05e06e07
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 Give the policy and the world model different jobs
The policy is trained to imitate successful action sequences, so its samples provide structured starting points for control. The predictive model has a different obligation: it must forecast what will happen when a candidate action is executed, including actions that depart from expert behavior. GPC therefore augments the world-model dataset with random perturbations while preserving expert-only policy training. Appendix A makes the asymmetry concrete: simulated Push-T includes 500 expert demonstrations and 3000 perturbation trajectories, plus seven long-horizon demonstrations. Figure 5 then compares planning with and without exploration. The reader's interpretation is that corrective action selection requires predictive coverage beyond the imitation manifold; the experiment supports the combined recipe but does not disentangle extra data volume from extra behavioral diversity. e04e05e13e19
Figure 2. Two nested processes build a visual rollout: denoise one next image, then advance the physical prediction. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left column as progression through physical time: observation history and action feed a conditional predictor, whose next image becomes part of the history for subsequent predictions. The right panel expands a single prediction. History enters CONV IN, the action enters the UNet, and CONV OUT contributes to the less noisy image. The displayed superscripts decrease from tau to tau minus one, consistent with Equation (7); they are diffusion indices rather than robot time. The caption specifies repeated denoising. Section IV adds a deployment detail absent from the drawing: the initial noise is fixed to zero, with three denoising steps. e07e08e21
What it supports. This design exposes predicted images to a reward model or VLM, making candidate consequences directly inspectable. Training moves from one-step supervision to supervision of recursive sequences. The frozen-noise rule makes action comparisons deterministic and is intended to stabilize the gradients used by OPT; it does not establish prediction accuracy by itself.
Where the evidence stops. The notation needs care: Section IV says T recursive applications produce an inclusive T+1-frame sequence, while Section III reports a prediction horizon of 16. The crop preserves the original labels; exact implementation lengths cannot be resolved from them.
5.2 Follow gradients through predicted outcomes, then execute actions
GPC-RANK only needs to compare candidate outcomes, which permits a nondifferentiable VLM to act as a selector. GPC-OPT instead differentiates a numerical reward through the world model and adjusts the input action chunk. Equation (4) uses positive-step gradient ascent, so an improving optimization trace means increasing predicted reward. The policy's weights are not updated. For visual prediction, denoising happens inside each physical prediction step, and recursively generated frames form the imagined future. Section IV fixes the initial diffusion noise to zero at inference to remove variation unrelated to the candidate actions. After selection or refinement, control executes an action segment and replans from new observations. The resulting distinction matters: optimized predicted reward, accurate imagined images, and successful execution are three separate quantities. e03e04e06e07e08e11
5.3 Evaluate improvement under the budget actually available
Table III makes inference effort part of the method rather than a secondary implementation concern. Combining 10 proposals with 25 refinement steps obtains the highest listed visual Push-T IoU, yet requires hundreds of seconds for a decision in that setting. A reader should therefore ask which planner is best under an available latency budget, not just which column has the highest score. Table II addresses another question: how GPC compares with other enhancements across four simulated tasks, using its separate 50-seed protocol. Figure 8 adds real execution, but only over 10 trials per reported evaluation. My assessment is that these results support the value of predictive selection while leaving fast closed-loop deployment, statistical robustness, and transfer to other task distributions incompletely established. e10e12e15e16e18
5.4 Training and inference
During training
Expert demonstrations are segmented with sliding windows. The policy uses DDPM, a UNet, and a ResNet18 visual encoder; training lasts 300 epochs with AdamW, learning rate 10^-4 and weight decay 10^-6. Images are resized to height 96. e04
Only the world-model dataset adds random perturbations to expert trajectories. Visual training first supervises one-step prediction, then jointly supervises recursively predicted sequences. The model uses EDM, observation horizon H=4 and three denoising steps. e05e08
A ResNet18+MLP predicts numerical rewards from images. Real Push-T reward labels use AprilTag-derived registration losses; cloth labels use normalized progress through successful grasp-and-move operations. State information used to label rewards is distinct from the visual-only real-world inference inputs. e11e15
During inference
The policy uses 100 denoising steps, reported observation horizon H=2, prediction horizon T=16 and action horizon 9. Control replans after executing an action segment. The paper's inclusive indexing is inconsistent with its horizon prose, so exact tensor lengths remain unresolved. e04e21
The world model initializes diffusion noise to zero at inference. A differentiable reward enables OPT; RANK can instead query a VLM once per decision using each candidate's final predicted frame and a task prompt, with temperature 0.2. The reported VLM setting uses 10 candidates versus 100 for learned-reward comparisons. e08e10e11
5.5 Implementation flow
- Propose from an expert prior
The stochastic diffusion policy maps observation history to plausible action chunks. Its parameters stay frozen during deployment; optimization changes the proposed actions. e03e04e06
- Predict candidate consequences
State-based dynamics use MLPs. Visual dynamics recursively predict the next image from observation history and the current action, feeding predictions into subsequent steps. e07e09
- Score and choose
GPC-RANK chooses the highest predicted reward among K proposals. GPC-OPT warm-starts from one proposal and performs M gradient-ascent steps through the reward and world model. Their combination optimizes multiple initializations before selection. e03e06
6. Experiments & results
Generative Predictive Control (GPC) improves a frozen diffusion policy by predicting the consequences of its action proposals, then selecting or refining them before execution. Its separate world model learns from demonstrations plus random exploration. Simulation and small real-world studies support this combination, while rollout cost and incomplete evaluation details limit the strength of broader deployment claims (e02, e03, e10, e12, e16, e18).
6.1 Read the original evidence
Table II. Learned-reward GPC-RANK has the highest reported score in each simulation task. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across one row at a time: each row is a different manipulation task, so the numbers do not define a single shared success scale. The rightmost unparenthesized entry uses GPC-RANK with a learned reward predictor and 100 proposals. Parentheses instead identify the VLM selector using 10 proposals; they are not error bars. Section V-B states that the reported scores average 50 evaluation seeds and that the enhancement baselines share training data and a pretrained diffusion policy. Keep that protocol separate from Table III, whose Push-T score and timing ablation uses 100 seeds, even where numerical entries coincide. e10e12
What it supports. GPC-RANK reaches 0.739 on visual Push-T versus 0.642 for Diffusion Policy and 0.683 for LaDi-WM. On Block Stacking, its 0.989 score exceeds DreamerV3's 0.983 by only 0.006. The ranking favors GPC across rows, but the magnitude of improvement varies substantially by task and comparator.
Where the evidence stops. Table II does not provide uncertainty estimates or explicit metric formulas for the non-Push-T tasks. The learned-reward and VLM columns also differ in candidate count, so their comparison does not isolate reward-model quality.
Figure 8. Physical trials show gains on pushing and folding, with limited statistical resolution. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the first three bars as a comparison within real-world Push-T, then read the last two as a separate clothes-folding comparison. The y-axis is success rate from zero to one; these values represent task execution rather than similarity between predicted images. Section V-C supplies the sample count missing from the chart: 10 trials for each reported evaluation. Figures 6 and 7 show representative trajectories, and their captions identify RANK with 10 proposals and no refinement. For Push-T the baseline bar reaches 0.5 and both GPC bars reach 0.7; for folding, baseline and RANK reach 0.3 and 0.7. e15e16e17e03e20
What it supports. The physical results extend the evidence beyond simulation and imagined videos. They suggest that predictive selection helps both rigid-object manipulation and a deformable-object task in the tested setup. Folding improves by 0.4 in plotted success rate, while the two Push-T variants each improve by 0.2 over baseline.
Where the evidence stops. Ten trials and no intervals provide limited precision; operational success thresholds are not specified. No cloth-OPT bar is shown. Figure 6 labels Push-T OPT as K=0,M=25, conflicting with Algorithm 1's K=1 warm start; that real-world configuration remains unresolved.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| State-based Push-T Simulation; mean over 100 evaluation seeds, Table I. | GPC-RANK: 0.934 with K=5000, M=0. Mean intersection-over-union (IoU), higher is better | Behavior cloning: 0.812; planning with a ground-truth simulator: 0.952. The best listed learned-model variant approaches the simulator reference; this state-input result is separate from vision-based evaluation. e09 |
| Vision-based simulation task comparison Table II; 50 evaluation seeds. Authors report common training data, the same pretrained diffusion policy and 100 candidates for enhancement baselines; VLM uses 10. | GPC-RANK: Push-T 0.739; Triangle Drawing 0.767; Block Stacking 0.989; Cube & Sphere Swapping 0.730. Reported task scores; non-Push-T score definitions are not supplied explicitly | Diffusion Policy: 0.642, 0.724, 0.912, 0.680. VLM variants: Triangle 0.761, Stacking 0.973. Learned-reward GPC leads every row. VLM numbers use a different candidate budget; scores should not be averaged across tasks or relabeled success probabilities. e10 |
| Vision-based Push-T planning–latency tradeoff Table III; simulation IoU over 100 evaluation seeds; wall-clock seconds per decision. | RANK (K=100,M=0): 0.739 / 11.745; OPT (K=1,M=25): 0.791 / 39.061; RANK+OPT (K=10,M=25): 0.882 / 374.102. Mean IoU / seconds per decision | Behavior cloning: 0.642 / 0.457; RANK (K=50,M=0): 0.698 / 5.835. Combined planning gains 0.240 absolute IoU over BC at substantial cost. The table labels BC as K=1,M=1, conflicting with the no-refinement baseline elsewhere. e12e03e09 |
| Random-exploration ablation Vision-based Push-T, Figure 5; world models trained with versus without exploration. | Adding exploration improves both RANK and OPT; no exact bar labels are supplied. Plotted score, higher is better | Expert-only world-model training gives lower bars for both planners. The trend supports broader dynamics coverage, but the comparison adds data and does not isolate exploration diversity from dataset size. e13e05 |
| Real-world Push-T and clothes folding Figure 8 and Section V-C; 10 trials for each reported evaluation. | Push-T RANK and OPT: 0.7 each; clothes-folding RANK: 0.7. Success rate read from plotted bars | Push-T baseline: 0.5; clothes-folding baseline: 0.3. These are executed-task outcomes, unlike imagined-image quality. Small trial counts, absent intervals and unspecified success thresholds limit precision. e15e16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table III. Read manipulation accuracy together with the time needed to make each decision. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the column headers to separate two computation budgets: K counts action proposals and M counts gradient updates. The first numerical row measures mean IoU across 100 evaluation seeds; the second reports wall-clock seconds for one decision cycle. Increasing the ranking budget from 50 to 100 improves the listed score but also increases latency. OPT uses one policy initialization with 25 refinement steps, whereas the combined column refines 10 initializations. Section III explains why these costs differ: candidate ranking can evaluate proposals in parallel, while refinement follows a sequence of dependent gradient updates. The table's actual wall-clock times remain the relevant deployment measurements. e12e03e06e09e18e20
What it supports. The combined setting reaches 0.882 IoU versus 0.642 for behavior cloning, an absolute gain of 0.240. It takes 374.102 seconds per decision versus 0.457 seconds. Ranking with 100 candidates achieves a smaller score of 0.739 at 11.745 seconds, illustrating a materially different accuracy–latency choice.
Where the evidence stops. The BC header prints M=1, conflicting with Table I's M=0 and the algorithm's meaning of refinement. The faithful crop retains that discrepancy. Hardware is unspecified, and these simulation timings should not be conflated with the roughly three-second real-world RANK setting.
Figure 5. Both planning strategies benefit when the world model also sees random perturbations. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the blue and orange bars within each planner before comparing the planners themselves. Blue denotes world-model training without random exploration; orange includes exploration. The horizontal labels distinguish ranking from gradient refinement, and both groups show an increase with exploration. The vertical axis begins at 0.4 rather than zero, which visually magnifies differences in bar height. Section III explains the intervention: humans or other controllers perturb the system without needing to complete the task, and these trajectories augment expert demonstrations for world-model learning. The policy still learns its behavior prior from expert demonstrations. The chart reports no exact labels on the bars. e05e13
What it supports. The observed trend is consistent with the authors' explanation that a world model needs experience beyond successful demonstrations to support corrective planning. Both RANK and OPT benefit in vision-based Push-T. This supports the combined data recipe, while leaving the precise source of the improvement to be tested separately.
Where the evidence stops. Adding exploration also adds training data; no matched-volume control is reported here. The chart lacks uncertainty bars and detailed per-condition training budgets. Treat the authors' approximate percentage description as qualitative rather than an exact relative or percentage-point gain.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify inference cost as the main limitation: world-model rollouts consume approximately 90–95% of runtime. Their approximately 3-second real-world RANK cycle is a different setting from Table III's simulation timings. e18e12
High SSIM and plausible predicted interactions do not establish accurate reward gradients or reliable recovery outside the tested tasks. Zero noise makes predictions deterministic; it does not by itself prove the authors' 'most likely future' interpretation. e14e08
Reporting conflicts remain: Table III gives BC M=1 while Table I gives M=0; Figure 6 gives OPT K=0 while Algorithm 1 requires a policy warm start with K=1. Horizon indexing also needs clarification. e12e09e17e03e21
7.2 Questions for discussion
- At matched wall-clock budgets, when does gradient refinement outperform ranking alone?
- Does exploration improve corrective dynamics because of its distribution, its volume, or both?
8. Reproducibility audit
8.1 Requirements and known gaps
Appendix A lists simulation expert/perturbation counts: Push-T 500/3000 plus seven long-horizon demonstrations; Triangle Drawing 30/100; Block Stacking 50/100; Cube & Sphere 100/100. Real tasks use 100 expert demonstrations per task and five random-play videos, each several minutes. e19
The supplied methods do not specify exact data splits, world-model phase durations, optimization step-size schedules, action constraints, GPU hardware or software versions. Task success thresholds and several score definitions are also absent. These gaps prevent exact reconstruction from this PDF alone. e20
Proposed checks should hold the policy fixed and test exploration at matched data volume, then compare zero, fixed-random and resampled world-model noise under matched planning budgets. Measure executed IoU and prediction–outcome agreement, not only predicted reward. e05e08e12e13
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 exploration diversity from extra training volume
Reader-proposed, not performed: freeze one Push-T diffusion policy and reward predictor, and train world models on expert-only data, expert-plus-exploration data, and an expert-only control with matched training frame count and optimizer updates. If additional independent demonstrations cannot be collected, explicitly label repeated expert frames as an exposure-matched rather than diversity-matched control. Use identical held-out initial states and candidate action chunks for ranking, with K=100 and zero inference noise. Report mean executed IoU over 100 evaluation seeds, paired uncertainty, and agreement between predicted and realized candidate rewards. The diversity explanation predicts that exploration improves corrective-candidate evaluation beyond the matched expert control; disappearance of that advantage would weaken that explanation. e05e08e12e13e19
Check 2: Test whether frozen noise improves real optimization rather than only its trace
Reader-proposed, not performed: for one trained visual world model, policy and reward predictor, compare zero initialization, a nonzero noise realization held fixed throughout each optimization, and fresh noise at each optimization evaluation. Hold starting policy samples, learning-rate schedule and M=25 constant, then repeat comparisons under matched measured latency against ranking. Evaluate actual simulator IoU as well as predicted reward before and after refinement, and report their disagreement across 100 common evaluation seeds. Fixed nonzero noise separates determinism from the special choice of zero noise. If predicted rewards rise without executed improvement, refinement is exploiting model or reward error; if fixed nonzero noise matches zero, determinism rather than a privileged zero-noise prediction explains the benefit. e06e08e09e12e20
8.3 Reading coverage
Visual audit: Visually inspected every supplied PDF page: p. 1 identity/version/affiliations; p. 2 Figure 1 and Algorithm 1; p. 3 policy training, horizons, datasets and Equations (1–5); p. 4 Figure 2, Equations (6–7), world-model training and frozen-noise remark; p. 5 Tables I–II and Figures 3–4; p. 6 Table III, Figure 5, reward and evaluation protocol; p. 7 real-world protocol and runtime limitation; p. 8 Figures 6–8 and Appendix A. All six final crops were opened and inspected at their original rendered resolution. Figure 1 selection/gradient directions and Figure 2 conditioning/denoising labels were checked against captions and equations. Table III's BC M label, Figure 6's OPT K label and horizon indexing conflicts are preserved and disclosed. The real-world bars were read against the axis; Figure 5 has no exact bar labels, so no precise values are inferred. Separate supplements, code and experiments remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. 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
- PDF pp. 1–2: title, abstract, I. Introduction, II. Related Work, Figure 1 and Algorithm 1
- PDF pp. 3–4: III. Overview of Generative Predictive Control; IV. World Model Learning
- PDF pp. 4–7: V. Experiments, including A. State-based Planar Pushing in Simulation, B. Vision-based Tasks in Simulation, C. Real-world Vision-based Tasks
- PDF p. 7: VI. Conclusion; VII. Limitation and Future Work
- PDF pp. 7–8: References
- PDF p. 8: Appendix A. Tasks and Datasets
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 all eight supplied PDF pages, all figures and tables, and six final original crops. All four supplied text chunks were read individually.
- Identity/version: the exact title and all five catalog authors match PDF p. 1. The reviewed artifact is arXiv:2502.00622v4 [cs.RO], 11 March 2026, an IEEE Robotics and Automation Letters accepted preprint; its footnote records revision on 31 January 2026 and acceptance on 28 February 2026. The catalog publicationYear is 2025 and its BibTeX describes a 2026 journal issue. Earlier versions and that final issue were not supplied or compared; their equivalence is not asserted (e01).
- No code, project website, external references, or separately hosted supplements were inspected. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, left-margin arXiv stamp and author/publication footnote
Exact title and authors Han Qi, Haocheng Yin, Aris Zhu, Yilun Du, Heng Yang; arXiv:2502.00622v4, 11 March 2026. Accepted RA-L preprint; revised 31 January and accepted 28 February 2026. Harvard University and Georgia Institute of Technology affiliations are listed.
Go to primary source ↓e02PDF p. 1, Abstract and Section I, motivating question and Contribution
GPC adds predictive online planning to pretrained frozen diffusion policies to address behavior-cloning deployment errors without policy retraining.
Go to primary source ↓e03PDF p. 2, Figure 1(a–b), caption, Algorithm 1 lines 1–10 and instantiation comment; Section II modular-design paragraph
Policy and world model are independently trained. K proposals may receive M action-gradient updates before ranking. The comment defines RANK by M=0 and OPT by K=1,M>0. Figure 1 stars the highest illustrated reward and shows gradients returning toward the action proposal.
Go to primary source ↓e04PDF p. 3, Section III, Generative policy training, including implementation paragraph
Sliding-window demonstrations train a DDPM policy with UNet and ResNet18 vision encoder. Reported H=2,T=16, action horizon 9, image height 96, 300 epochs, AdamW learning rate 10^-4, weight decay 10^-6 and 100 inference denoising steps; receding-horizon control.
Go to primary source ↓e05PDF p. 3, Section III, Predictive world modeling
World-model training combines expert trajectories with human/controller random perturbations that need not solve the task, motivated by broader dynamics excitation.
Go to primary source ↓e06PDF p. 3, Section III, Online planning, Equations (1–5); PDF p. 4, opening GPC-RANK+OPT paragraph
K sampled action chunks are scored through R(W). OPT maximizes predicted reward by positive-step gradient ascent from a policy sample for M iterations. Ranking supports nondifferentiable rewards; the combined method refines multiple initializations.
Go to primary source ↓e07PDF p. 4, Figure 2 and caption; Section IV, Equations (6–7)
Visual dynamics recursively apply an action-conditioned one-step image predictor. Observation history enters CONV IN, action enters UNet, and denoising maps noise index tau to tau-1. MLP dynamics are available for low-dimensional states.
Go to primary source ↓e08PDF p. 4, Section IV, Two-phase training and Remark 1 (Freeze the Noise)
One-step supervision precedes recursive multistep supervision. H=4 and N_d=3 with EDM. Initial image noise is fixed to zero at inference; authors state stochastic gradients otherwise destabilize OPT and call the deterministic output the most likely prediction, without a formal mode argument.
Go to primary source ↓e09PDF p. 4, Section V-A; PDF p. 5, Table I and Section V-A continuation
State Push-T uses simulator states, MLP dynamics and an object-vertex registration loss. Table I averages IoU over 100 seeds: BC 0.812 (K=1,M=0), RANK 0.934 (K=5000,M=0), ground-truth simulator reference 0.952.
Go to primary source ↓e10PDF p. 5, Table II, all rows/columns and caption; PDF p. 6, Section V-B, Planning performance
Rows Push-T/Triangle/Stacking/Swapping: Diffusion Policy 0.642/0.724/0.912/0.680; V-GPS 0.620/0.593/0.972/0.650; LaDi-WM 0.683/0.680/0.933/0.410; DreamerV3 0.649/0.732/0.983/0.700; GPC 0.739/0.767/0.989/0.730. Parenthesized VLM scores are 0.761 and 0.973. Text specifies 50 seeds, common data and pretrained policy, 100 candidates for comparisons; caption distinguishes learned-reward K=100 from VLM K=10. Other task score formulas are not stated.
Go to primary source ↓e11PDF p. 6, Section V-B, Reward paragraph
Differentiable ResNet18+MLP reward predictors use numerical objectives such as registration loss or cube distance. Optional ChatGPT-4o selects from final predicted frames at resolution 96 with task-specific prompts, one query per decision and temperature 0.2.
Go to primary source ↓e12PDF p. 6, Table III, all configuration headers, Score (IoU)/Timing (s) rows and caption
Over 100 seeds, BC/RANK50/RANK100/OPT25/RANK10+OPT25 score 0.642/0.698/0.739/0.791/0.882, with 0.457/5.835/11.745/39.061/374.102 seconds per decision. The printed BC header says K=1,M=1, unlike Table I's M=0.
Go to primary source ↓e13PDF p. 6, Figure 5, legend, bars and score axis; Section V-B, Importance of random exploration
With-exploration bars exceed without-exploration bars for both RANK and OPT on vision-based Push-T. The axis starts at 0.4; exact bar values, uncertainty and a size-matched data control are not provided. Text describes an approximately 10% improvement without clarifying that percentage convention.
Go to primary source ↓e14PDF p. 5, Figures 3–4 and captions; PDF p. 6, Section V-B opening image-quality discussion
Predicted frames and SSIM assess visual correspondence. Figure 3 uses the full roughly 250-frame evaluation horizon over five seeds; Figure 4 uses ten sampled prediction points and 100 samples per method. These are distinct image-quality protocols, not executed-task success rates.
Go to primary source ↓e15PDF p. 7, Section V-C, Real-world Vision-based Tasks
Push-T reward supervision uses AprilTags and registration loss; cloth supervision uses normalized progress from spread clothes through grasp-and-move operations. All inference components use visual observations. Real-world success rates are evaluated over 10 trials.
Go to primary source ↓e16PDF p. 8, Figure 8, all five bars and Success Rate axis; PDF p. 7, Section V-C trial-count sentence
Axis-read success rates: Push-T baseline 0.5, RANK 0.7, OPT 0.7; clothes baseline 0.3, RANK 0.7. No error bars, confidence intervals or explicit operational success thresholds are supplied; Section V-C states 10 trials.
Go to primary source ↓e17PDF p. 8, Figures 6–7 and captions; PDF p. 2, Algorithm 1 instantiation comment
Real trajectories contrast BC and GPC. Figure 6 labels RANK K=10,M=0 and OPT K=0,M=25; the latter conflicts with Algorithm 1's K=1 OPT warm start. Figure 7 reports cloth RANK K=10,M=0 and no cloth OPT trajectory.
Go to primary source ↓e18PDF p. 7, Section VII, Limitation and Future Work
Authors attribute approximately 90–95% of runtime to diffusion world-model rollouts and report about three seconds per real-world RANK decision. Efficiency improvements are future work.
Go to primary source ↓e19PDF p. 8, Appendix A, Tasks and Datasets, complete paragraph and bullets
Simulation: Push-T 500 expert demos, 6×500 perturbations and seven long-horizon demos; Triangle 30/100, Stacking 50/100, Cube & Sphere 100/100 expert/perturbation trajectories. Real tasks: 100 expert demonstrations per task and five random-play videos of several minutes.
Go to primary source ↓e20PDF pp. 3–4, implementation/training and optimization descriptions; pp. 5–7, evaluation descriptions; p. 8, Appendix A
These supplied implementation sections give selected policy hyperparameters, world-model horizons and data totals, but omit exact splits, phase schedules, action constraints, optimization step-size values, GPU/software specifications and operational real-world success thresholds.
Go to primary source ↓e21PDF p. 3, Section III opening notation and implementation paragraph; PDF p. 4, Figure 2 and Two-phase training
The notation defines a_t:t+T as T+1 actions and inclusive histories, while implementation calls T=16 a prediction horizon and H=2 an observation horizon. Training prose says T applications yield o_t+1:t+T+1, an inclusive T+1-frame range. Exact indexing is unresolved.
Go to primary source ↓8.5 Primary sources
Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling ↗
PDF · 5,777 extracted words
Source fingerprint
53f1ed78305b3f19fcfd6dfd13e942241d19ebc3cf531e93a38138af791b5962