LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE
1. Paper overview
In one sentence: LongScape turns action-derived chunk lengths into a choice among four video denoisers, improving reported long-horizon video quality while leaving routing causality and execution utility untested. identitypartitionexpertsroutermain-results
| At a glance | What to know |
|---|---|
| Research problem | Author claim Fixed temporal chunks can divide a manipulation event or mix different motion regimes. The authors argue that this undermines long-horizon coherence, while discrete autoregressive generation sacrifices visual detail. Their target is a video generator that combines local diffusion fidelity with sequential generation and context-dependent temporal granularity. motivationfactorization |
| Core mechanism | Source description A preprocessing heuristic uses gripper changes and six-dimensional motion ranges to construct 8-, 16-, 24- and 32-frame chunks. partition |
| A key reported result | LIBERO long-horizon video generation: 19.977 / 0.1231 / 0.7883 / 153.72 PSNR / LPIPS / SSIM / FVD. Approximately 20-second composite manipulation videos; exact evaluation split and video count unspecified. CogVideoX generates the whole sequence, while autoregressive/hybrid models roll out iteratively. CogVideoX: 19.315 / 0.1402 / 0.7729 / 184.69, the strongest listed baseline in every LIBERO column. All four metrics favor LongScape. These are video-quality results without reported uncertainty, not manipulation success rates. main-resultscomparisonsetup |
| Reading caution | Source description The authors identify heuristic chunking as a limitation and leave learned partitioning, multiple camera views and several-minute generation to future work. future-work |
Core contributions
- Source description
A preprocessing heuristic uses gripper changes and six-dimensional motion ranges to construct 8-, 16-, 24- and 32-frame chunks. partition
Figure 1. Action annotations determine the training video's temporal units. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each purple boundary as one final chunk, then connect its length above to the action cue below. The eight-frame example isolates an effector-state change; the other examples terminate when an arm's accumulated action range crosses a threshold. Section 3.2 and Algorithm 1 explain how these lengths arise: start with eight frames and grow the candidate until an ordered stopping rule applies. A threshold crossing can therefore occur at 32 frames as well as at 16 or 24. The figure illustrates boundary construction, rather than a guarantee that every large-motion event always produces the shortest chunk. partitionalgorithmrouter
What it supports. Temporal granularity is supervised by recorded robot behavior before the generator is trained. The four allowed durations become four expert classes. This connects the segmentation heuristic to the learned routing problem, but it does not make the resulting units verified semantic action labels or equip inference with future robot commands.
Where the evidence stops. The threshold fraction alpha is unspecified. Motion detection and the maximum-length branch precede gripper splitting in the written algorithm, so the caption's semantic-unit description is an intended effect, not an unconditional guarantee for every event boundary.
2. Motivation
2.1 The problem and the proposed response
Fixed temporal chunks can divide a manipulation event or mix different motion regimes. The authors argue that this undermines long-horizon coherence, while discrete autoregressive generation sacrifices visual detail. Their target is a video generator that combines local diffusion fidelity with sequential generation and context-dependent temporal granularity. motivationfactorization
2.2 What this reading follows
A robot approaching an object and closing its gripper need different kinds of temporal detail. LongScape uses that observation to decide the size of the next video segment. During preprocessing, recorded actions determine chunk boundaries. During generation, a learned router replaces those action annotations with a decision based on the instruction and recent visual context. Four diffusion experts then generate different chunk lengths, and their outputs become the context for later predictions. This edition follows the supplied September 2025 v1 from partition rules through training and inference to its two quantitative tables. The central evidence concerns generated-video quality; the architecture does not output a robot policy. identitypartitionexpertsroutermain-results
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 recorded Foundational work / Neural world simulators placement is supported: LongScape learns visual dynamics and generates videos. Its action annotations shape training chunks, while inference predicts future visuals from text and context. No action head, inverse-dynamics action extraction or policy execution is specified, so the action-model architecture/prediction quadrants remain not applicable. Four generative experts do not establish a joint future-and-action One Model architecture. partitionexpertsrouter
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 action traces into a temporal vocabulary
LongScape first uses information that generation will not have: the recorded robot's gripper and motion trace. It partitions eight-frame base units into four allowed lengths, then treats length as the expert label. To understand the heuristic, imagine a candidate whose motion remains below threshold and whose gripper state does not change. It keeps growing until another rule fires or the four-unit cap is reached. A candidate crossing the motion threshold is committed immediately, with that crossing included. A gripper change splits off the final unit only if the earlier branches have not already fired. This is why rule precedence and missing edge-case handling matter. The semantic interpretation is an author motivation; the actual training labels come from a deterministic action-range heuristic. partitionalgorithmmotivation
Figure 2. Generating a chunk and selecting its generator have different training targets. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the lower half first: eight-, sixteen-, twenty-four- and thirty-two-frame examples align with Experts 1 through 4. The diffusion-loss links refer to Equation (2), where an expert predicts noise in a target latent using the instruction and preceding visual context. Now move upward: the router predicts expert identifiers, shown as predicted and ground-truth labels joined by cross-entropy losses. Equation (3) trains this classifier on instruction, current chunk and next-expert label. The diagram summarizes these two learning tasks; it does not depict the noisy latent, diffusion time or every preceding-context connection needed by the equations. expertsrouterpartitionsetup
What it supports. Specialization operates at the level of four CogVideoX-initialized denoisers. The router learns when to use them instead of synthesizing pixels itself. Its correctness is defined by preprocessing labels, which means excellent label prediction can coexist with imperfect chunk boundaries or imperfect generated motion.
Where the evidence stops. The figure does not establish shared weights, encoder freezing or a joint end-to-end training schedule. The paper gives separate objectives and training times, but omits enough optimizer and configuration detail that this graphic alone cannot specify a reproducible training recipe.
5.2 Separate predicting noise from predicting duration
Once the chunks exist, there are two supervised problems. Each DiT expert learns to remove noise from a target chunk of its assigned length, conditioned on text and preceding visual context. The router instead learns a classification target: which expert should generate the next chunk? At inference, the router's argmax determines the temporal extent of the next diffusion problem. The chosen expert produces pixels through latent denoising, and generated context then feeds the next decision. Reader analysis: this separates local image synthesis from temporal scheduling, but it also couples their failures over time. A poor generated frame can distort routing even if the classifier performs well on its original test samples. Neither the causal factorization nor the router's label accuracy guarantees physical consistency. factorizationexpertsrouterrouter-test
Figure 3. Generated visual context determines which duration to predict next. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the initial frame and instruction on the left. Their encoded information reaches the router, which selects a generator for the next chunk. The green dotted arrows carry generated context into subsequent routing decisions; the text specifies last-frame visual features. Follow the chunks down to concatenation and VAE decoding to see the final output: a video. For the decision rule, use Equations (4)–(5): select the maximal router score and invoke that expert's iterative denoising process. The picture's score bars lack class labels, so their example destinations should not override this explicit mathematical rule. routerexperts
What it supports. LongScape adapts its video horizon one chunk at a time without requiring a future action sequence at inference. This also makes routing vulnerable to its own generated context: a visual error can influence the next duration decision. The diagram establishes this feedback path, not a measured robustness result.
Where the evidence stops. If bars follow expert order, the middle group's third-bar maximum points to Expert 2 and the right group's fourth-bar maximum points to Expert 3. This appears inconsistent with argmax routing. The faithful figure is retained; its unlabeled bar mapping remains unresolved, while the report follows Equations (4)–(5).
5.3 Ask what each experiment can identify
Table 1 answers whether the complete LongScape system scores better than the listed baselines on generated-video metrics. Table 2 narrows the comparison to fixed-length single experts and shows that none matches the full adaptive system. Figure 7 then checks how often the router reproduces its assigned expert label. These are complementary tests, but they do not form a clean causal decomposition. Reader analysis: to credit semantic partitioning specifically, a control needs similar capacity and training exposure with boundaries unrelated to actions. To credit the router specifically, the experts should stay fixed while routing changes. The qualitative frames help locate plausible failure modes, but sampled images and labels such as successful action are not measured robot execution. This distinction limits how far the system's data-generation promise has been tested. main-resultsablationrouter-testqualitativequalitative-ablationexperts
5.4 Training and inference
During training
Videos are sampled at 10 Hz, yielding approximately 30,000 training clips from each dataset. Each expert has 5.57G parameters; routers have 71.35M parameters for LIBERO and 108.57M for AGIBOT-World. Each expert trains for two epochs, reportedly about 24 hours on four NVIDIA H20 GPUs. Router training separately takes about 24 hours on four H20s. setup
During inference
Encode the initial image and instruction. At each step, the router uses the instruction and visual features from the current chunk's last frame, chooses the highest-scoring expert, and lets that expert denoise the next chunk. Concatenate generated chunks and decode the video. The feedback is generated visual context; the described pipeline contains no physical execution or environmental observation loop. router
Figure 3 leaves its score-bar class order unlabeled. Under the natural left-to-right expert ordering, its middle and right maxima do not match the arrows to Experts 2 and 3. Equations (4)–(5) unambiguously specify argmax routing; the illustrative assignments remain unresolved. router
5.5 Implementation flow
- Partition demonstrations
Begin with non-overlapping eight-frame units. For each positional dimension, set a threshold to a fraction alpha of its whole-video motion amplitude. Grow a candidate from one to four units. Commit it when any motion range exceeds its threshold; otherwise commit at four units; otherwise isolate the final unit if its gripper state changes; otherwise expand. This branch order matters: motion and maximum length take precedence over gripper splitting. partitionalgorithm
- Specialize denoisers
Initialize each of four experts from pretrained CogVideoX parameters. Each uses transformer blocks with 3D full attention, feed-forward networks and AdaLN, and predicts noise for its assigned chunk length conditioned on text and preceding video context. This is routing among full generators, rather than token-level switching inside one shared DiT. experts
- Learn the next chunk class
A single cross-attention transformer predicts a ground-truth expert index derived from partition lengths. Its cross-entropy objective is distinct from the experts' noise-prediction objective; the paper does not specify an end-to-end joint optimization schedule. expertsrouter
6. Experiments & results
LongScape generates embodied manipulation videos by choosing how many frames to predict next. Robot action annotations define variable-length training chunks; four diffusion experts specialize in their lengths, and a text-and-vision router selects an expert during autoregressive rollout. The reported gains concern video similarity and distributional quality on LIBERO and AGIBOT-World. The method supplies neither executed robot actions nor evidence that its videos improve downstream policies.
6.1 Read the original evidence
Table 1. LongScape leads the reported video-quality metrics on both datasets. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Keep the two dataset blocks separate and follow each metric's direction arrow. PSNR and SSIM increase with the reported image similarity, while LPIPS and FVD decrease. On LIBERO, CogVideoX is the strongest baseline in all four columns, so one row gives a consistent comparison. On AGIBOT-World, the strongest comparator changes: Genie leads baseline PSNR and LPIPS, NOVA leads SSIM, and CogVideoX leads FVD. Section 4.2 describes approximately twenty-second composite tasks, with whole-sequence generation for CogVideoX and iterative rollout for autoregressive or hybrid methods. Thus the table compares complete systems under different generation structures. main-resultscomparisonsetupbaselines
What it supports. LIBERO FVD falls from CogVideoX's 184.69 to 153.72, alongside improvements in all three image metrics. AGIBOT-World FVD falls from 267.39 to 256.16; its LPIPS margin over Genie is only 0.0007. The consistent directions support a video-quality advantage, while the varying margins discourage treating every gain as equally substantial.
Where the evidence stops. The table contains no uncertainty, and the generation-test split and sample count are unspecified. It evaluates images and videos, not robot-executed success or downstream policy learning. Appendix A.1 also leaves baseline configurations too incomplete to establish a compute-matched comparison.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO long-horizon video generation Approximately 20-second composite manipulation videos; exact evaluation split and video count unspecified. CogVideoX generates the whole sequence, while autoregressive/hybrid models roll out iteratively. | 19.977 / 0.1231 / 0.7883 / 153.72 PSNR / LPIPS / SSIM / FVD | CogVideoX: 19.315 / 0.1402 / 0.7729 / 184.69, the strongest listed baseline in every LIBERO column. All four metrics favor LongScape. These are video-quality results without reported uncertainty, not manipulation success rates. main-resultscomparisonsetup |
| AGIBOT-World long-horizon video generation Approximately 20-second composite tasks from real-world robot video; exact evaluation split and count unspecified. | 16.493 / 0.3613 / 0.7015 / 256.16 PSNR / LPIPS / SSIM / FVD | Best baseline by column: Genie PSNR 16.166 and LPIPS 0.3620; NOVA SSIM 0.6998; CogVideoX FVD 267.39. LongScape leads each column, but the LPIPS and SSIM margins are small and their statistical reliability is unreported. main-resultscomparison |
| Adaptive experts versus fixed chunk lengths Single-expert variants trained exclusively on 8-, 16-, 24- or 32-frame chunks, evaluated on both datasets. | LongScape: 153.72 on LIBERO; 256.16 on AGIBOT-World FVD, lower is better | Best fixed-length FVD: LIBERO 16-frame model, 176.55; AGIBOT-World 32-frame model, 282.38. Adaptive LongScape wins, but performance is not monotonic in fixed length. The comparison changes expert count, specialization and partitioning together, so it cannot assign the gain uniquely to routing. ablationexperts |
| LIBERO next-expert classification Four-class router test on 12,700 samples. | 11,606 correct; reported accuracy 91.4% Expert-index accuracy | No majority-class, random-router or oracle-generation baseline is reported. The router usually reproduces heuristic labels. The test does not establish accuracy on drifting generated contexts or downstream video quality conditional on a routing error. router-testrouter |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. Adaptive generation beats each fixed-length variant, with no universal best fixed length. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down one metric column before comparing datasets. The four lower rows are single experts trained exclusively on their respective fixed lengths; the top row combines four specialized experts with adaptive routing. LIBERO's best fixed FVD is 176.55 for sixteen frames, even though the thirty-two-frame model leads the fixed variants in the other three LIBERO metrics. On AGIBOT-World, thirty-two frames gives the best fixed FVD, while other metrics prefer different lengths. These crossings matter: the evidence supports an adaptive system over the listed fixed systems, rather than a simple rule that longer chunks are always better. ablationexpertspartition
What it supports. LongScape's FVD is 153.72 on LIBERO and 256.16 on AGIBOT-World, outperforming the best fixed alternatives at 176.55 and 282.38. The pattern is consistent with different temporal scales being useful, but it does not determine whether semantic boundaries, expert specialization or additional total capacity contributes most.
Where the evidence stops. Replacing four routed specialists with one fixed-length expert changes capacity, training allocation and boundary structure together. There is no capacity-matched nonsemantic partition control or router-only ablation here, so a causal claim about routing alone would exceed this table.
Figure 7. The router usually matches its heuristic target, but error consequences are unmeasured. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the large bar at zero, then inspect the much smaller off-center bars. The vertical axis counts predictions; the horizontal ticks run from minus twenty-four to twenty-four in increments of eight. These increments are compatible with the spacing between allowed chunk lengths, but the paper does not define the sign convention or an explicit error formula. Section 4.3 provides the reliable aggregate: 11,606 correct assignments out of 12,700 LIBERO samples, reported as 91.4% accuracy. This plot is an aggregate error distribution, not a confusion matrix showing which true expert class is hardest to predict. router-testrouterpartition
What it supports. Text and visual context contain enough information for the router to reproduce most action-derived length labels in the stated test. This supports replacing unavailable future action annotations with a learned selection rule. It does not show that the heuristic targets are optimal or that every misclassification has the same effect on video quality.
Where the evidence stops. Per-class sample counts, a majority-class baseline, sign convention and an evaluation on generated contexts are absent. A high aggregate classification score therefore cannot by itself establish robust long-horizon routing or measure how a wrong duration affects subsequent chunks.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify heuristic chunking as a limitation and leave learned partitioning, multiple camera views and several-minute generation to future work. future-work
The introduction claims stability over 15 rollouts, but the paper supplies no horizon-by-horizon quantitative curve. Selected qualitative frames support examples of coherence; they cannot establish continuous physical validity, executed task success or downstream policy improvement. motivationqualitativequalitative-ablationmain-results
The stated average improvements of 8.6% and 5.8% lack an explicit aggregation definition. Different baseline sizes and training budgets are insufficiently documented for a compute-matched comparison. Activating one expert also does not establish peak-memory or latency parity; those measurements and weight-loading details are absent. comparisonbaselinesroutersetup
7.2 Questions for discussion
- Would a capacity-matched model with nonsemantic variable lengths retain the gains?
- How does an incorrect chunk-length decision affect subsequent generated contexts and router errors?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs aligned action/video data, instructions, the CogVideoX initialization, all four expert checkpoints and the contextual router. Appendix A.1 identifies the CogVideoX baseline as 5B, but exact Genie/NOVA configurations and training schedules are not supplied. partitionexpertsrouterbaselines
The numerical alpha threshold, optimizer, learning rate, batch size, resolution, sampling/guidance settings, frozen-module policy, software versions, seeds and generation-test split are unspecified. Algorithm 1 also omits explicit pointer advancement/reset, tail handling and treatment of the empty prefix when a gripper change occurs at n=1. These need clarification or declared implementation choices. partitionalgorithmexpertssetupcomparison
Proposed checks should separate semantic partitioning from extra capacity, and test learned routing under generated-context drift. Keep evaluation videos and budgets fixed, report uncertainty and compare against controlled partition or routing alternatives; neither check has been run here. ablationrouter-testrouter
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 action-aligned segmentation matter at matched capacity?
Reader-proposed, not run: compare the published action-guided partition with boundaries randomized within each training video while preserving its chunk-length histogram. Train four experts and a router in both arms, matching initialization, frame exposure and optimization budget. Evaluate identical held-out videos and prompts across seeds, reporting FVD and image metrics plus separately annotated boundary discontinuities. Specify alpha and algorithm edge cases before training. If the action-guided advantage disappears at matched capacity and length distribution, Table 2's gains would not uniquely support semantic boundary placement; a reproducible advantage would support that contribution. partitionalgorithmexpertsablationmain-results
Check 2: Does generated context degrade routing and amplify errors?
Reader-proposed, not run: freeze the trained experts and router. At aligned positions in held-out demonstrations, score next-length classification using either the ground-truth preceding image or a generated preceding image. Report per-class accuracy and confusion, then compare rollouts under learned routing, heuristic labels derived from the held-out action trace as an evaluation oracle, and a frequency-matched random router. Match final video duration, initial images, prompts and sampling settings; report FVD, boundary discontinuities and error versus elapsed time. A generated-context accuracy drop accompanied by an oracle rollout advantage would identify routing drift as a remaining bottleneck. routerrouter-testpartitionmain-resultsablation
8.3 Reading coverage
Visual audit: Visually inspected the title/authors/version page, all eight original figures, Tables 1–2, method equations, the implementation and evaluation passages, the conclusion's future-work paragraph, Appendix A.1 and Algorithm 1. All six final original crops were separately viewed; Figure 7 was cropped from a 400-DPI render, and the other assets from 200-DPI renders. Figure 3's score-bar/arrow ambiguity was cross-checked against Equations (4)–(5) and disclosed. Figure 1 was checked against the ordered rules in Section 3.2 and Algorithm 1. All four text chunks, including references and the appendix, were read. Reference-only p. 11 was read as text and is outside the page-image pass. No separate supplement, code or continuous video playback was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13. 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, authors, affiliations and arXiv:2509.21790v1 stamp, 26 September 2025 (p. 1)
- Abstract and Section 1 Introduction (pp. 1–2)
- Section 2 Related Works, including 2.1 and 2.2 (pp. 2–3)
- Section 3 Methodology, including 3.1–3.3 and Equations (1)–(5) (pp. 3–5)
- Section 4 Experiments, including 4.1–4.3 (pp. 6–9)
- Section 5 Conclusion and Future Works (pp. 9–10)
- References (pp. 10–11)
- Appendix A.1, Algorithm 1 and Figures 7–8 (pp. 12–13)
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 every figure, both tables, the equations and Algorithm 1 in the supplied PDF. References on p. 11 were read as text, without a separate page-image inspection.
- Only the supplied v1 was reviewed; no later revision or separate supplement was supplied or compared. Its title and all seven authors agree with the catalog; no identity discrepancy was found.
- The abstract states that code is available, but the linked repository was not inspected. No paper scripts were executed and no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title block, affiliations, abstract and arXiv margin stamp
Title matches the catalog. Authors are Yu Shang, Lei Jin, Yiding Ma, Xin Zhang, Chen Gao, Wei Wu and Yong Li; affiliations are Tsinghua University and Manifold AI. The stamp identifies arXiv:2509.21790v1, 26 Sep 2025. The abstract states code availability via a repository link.
Go to primary source ↓motivationPDF pp. 1–2, Abstract and Section 1
The authors motivate adaptive chunks by temporal drift and fixed-chunk semantic fragmentation, and claim stable generation over 15 rollouts.
Go to primary source ↓factorizationPDF pp. 3–4, Section 3.1, Equation (1) and conditional denoising discussion
Video is factorized into causal chunks; next-chunk generation uses continuous latent diffusion conditioned on preceding video context.
Go to primary source ↓partitionPDF p. 3, Figure 1; p. 4, Section 3.2
Action ranges and gripper changes guide grouping of eight-frame units into four possible lengths. Thresholds are a fraction alpha of per-video motion amplitude; no numerical alpha is given. Figure 1 labels both shorter and 32-frame examples with threshold crossings.
Go to primary source ↓algorithmPDF p. 12, Algorithm 1, lines 1–16; p. 4, Section 3.2 numbered rules
The ordered branches test motion above threshold, then n=4, then a gripper change, then expansion. The pseudocode does not explicitly advance/reset its pointer or specify empty-prefix and residual-tail cases.
Go to primary source ↓expertsPDF p. 5, Figure 2 and Section 3.3, Equation (2)
Four CogVideoX-initialized DiT experts specialize in chunk lengths, with 3D full attention and AdaLN; their conditional noise-prediction losses are separate from the illustrated router classification losses.
Go to primary source ↓routerPDF p. 5, Section 3.3, Equations (3)–(5); p. 6, Figure 3 and caption
The cross-attention router learns expert labels with cross-entropy and uses instruction plus current visual context, specifically last-frame features at inference. Argmax selects one expert. Figure 3 shows VAE/text encoding, generated-context feedback, concatenation and decoding; its unlabeled middle/right score maxima visually differ from arrow destinations under left-to-right expert ordering. Memory economy is asserted without measurements.
Go to primary source ↓setupPDF pp. 6–7, Section 4.1, Datasets, Metrics and Implementation of LongScape
Both datasets are sampled at 10 Hz with about 30,000 training clips each. Experts have 5.57G parameters; AGIBOT/LIBERO routers have 108.57M/71.35M. Experts train for two epochs, about 24 hours on four H20s; router training takes about 24 hours on four H20s. Metrics assess generated images/videos.
Go to primary source ↓comparisonPDF p. 7, Section 4.2, Quantitative comparison
Tasks last approximately 20 seconds. CogVideoX generates entire sequences while autoregressive/hybrid baselines use rollouts. Average improvements of 8.6% on LIBERO and 5.8% on AGIBOT-World are asserted without an explicit aggregation formula.
Go to primary source ↓main-resultsPDF p. 8, Table 1, all rows and both dataset column groups
LongScape reports LIBERO PSNR/LPIPS/SSIM/FVD 19.977/0.1231/0.7883/153.72 and AGIBOT-World 16.493/0.3613/0.7015/256.16. Table arrows favor higher PSNR/SSIM and lower LPIPS/FVD; each LongScape entry beats all listed baselines. No uncertainty is shown.
Go to primary source ↓qualitativePDF p. 7, Figure 4 and Section 4.2; p. 8, Figure 5 and accompanying discussion
Selected frames compare a LIBERO drawer task and AGIBOT cloth folding. The text attributes ghosting, incorrect object choice and disappearance to baselines. These are generated-video examples, not physical deployment trials.
Go to primary source ↓ablationPDF pp. 8–9, Section 4.3; p. 9, Table 2, all fixed-length rows
Single experts trained on fixed 8/16/24/32-frame chunks are compared with LongScape. LIBERO fixed FVD values for 32/24/16/8 frames are 180.31/191.02/176.55/229.71; AGIBOT values are 282.38/303.75/373.50/387.49. LongScape achieves 153.72 and 256.16.
Go to primary source ↓router-testPDF p. 9, Section 4.3, Effectiveness of the dynamic router; p. 12, Figure 7
The LIBERO router correctly classifies 11,606 of 12,700 samples, reported as 91.4%. Figure 7 plots counts against prediction error at ticks from -24 to 24 in steps of eight; it supplies no sign convention, class-conditioned confusion matrix or generated-context evaluation.
Go to primary source ↓qualitative-ablationPDF p. 9, Figure 6; p. 13, Figure 8 and captions
AGIBOT and LIBERO frame grids compare adaptive LongScape with four fixed-length variants. Figure 6 annotates generated sequences as successful, locomotion-only, discontinuous or limited-motion examples; these are illustrative labels rather than measured execution rates.
Go to primary source ↓future-workPDF p. 10, continuation of Section 5 before References
The authors identify heuristic partition rules and propose learned chunking, multi-view inputs and several-minute generation as future directions.
Go to primary source ↓baselinesPDF p. 12, Appendix A.1, Details of Baselines
CogVideoX uses the 5B version. Genie is described as discrete-token autoregressive generation and NOVA as diffusion within autoregressive frame generation; their exact configurations and training budgets are not detailed.
Go to primary source ↓8.5 Primary sources
LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE ↗
PDF · 5,590 extracted words
Source fingerprint
44347deaad2224fd086c0de7a721d5ebc37c80df311e95ed8d1e4f63ffcf5d6e