GigaWorld-Policy: An Efficient Action-Centered World–Action Model
1. Paper overview
In one sentence: A shared causal Transformer uses future-video supervision to train a robot policy whose deployed actions do not require future-video generation. motivationarchitectureinferencelatencyreal-resultsevaluationdata-efficiencysimulationfuture-ablation
| At a glance | What to know |
|---|---|
| Research problem | Author claim Sparse action labels may encourage shallow policies, while requiring video rollouts can make world-model control slow. The authors seek dense visual supervision without making each robot command depend on a generated future. motivation |
| Core mechanism | Source description A single shared Transformer separates action prediction from forward visual dynamics through directed attention dependencies, enabling an action-only deployment path. factorizationarchitectureinference |
| A key reported result | Four real-world PiPER manipulation tasks: Average 0.83; cleaning 0.90, QR scanning 0.75, sweeping 0.75, bowl stacking 0.90. Reported SR, higher is better. 50 post-training demonstrations per task; 20 trials per task, up to five attempts per trial; graded real-world scoring. Motus average 0.76; pi0.5 0.69. Rounded averages differ by 7 and 14 percentage points. QR scanning ties Motus. This is not one-attempt binary success. evaluationreal-resultstasks |
| Reading caution | Reader analysis The paper does not isolate latency as the cause of better physical scores: models, priors and decoding differ together. Graded credit, repeated attempts and missing uncertainty limit comparisons with conventional binary success rates. evaluationlatencyreal-results |
Core contributions
- Source description
A single shared Transformer separates action prediction from forward visual dynamics through directed attention dependencies, enabling an action-only deployment path. factorizationarchitectureinference
- Source description
A web-video → embodied-video → target-robot curriculum combines video priors with instruction-conditioned policy specialization. Its benefits are evaluated through pre-training and data-efficiency ablations. curriculumpretraining-ablationdata-efficiency
Figure 3. One backbone learns both streams; the inference panel makes future-video decoding optional. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the training panel from left to right. The VAE supplies observation tokens T_o and future tokens T_f; the noise symbols identify the future and action streams being denoised. State T_s and action T_a enter through their encoder, while text T_l conditions the blocks. The legend distinguishes video pre-training/post-training paths from post-training-only paths. Now compare the right panel: observations, state, language and noisy actions remain, but the future-token arrow is dashed and labeled optional. The KV cache supports later video generation. Section 3.2 clarifies that text uses cross-attention, so its bottom arrow should not be read as another position in the causal self-attention sequence. factorizationarchitectureobjectivesinferencecurriculum
What it supports. The architecture allows visual supervision to shape shared parameters during learning while keeping future-video computation outside the required control path. The paper describes both joint optional video denoising and later video denoising from the cached action context; neither is required to produce the reported action-only policy output.
Where the evidence stops. The figure is a schematic, not a complete training specification. Eqs. (2)–(3) introduce an action conditioning latent, whereas the losses use action tokens; their exact implementation mapping and missing-action handling during video pre-training remain unspecified.
2. Motivation
2.1 The problem and the proposed response
Sparse action labels may encourage shallow policies, while requiring video rollouts can make world-model control slow. The authors seek dense visual supervision without making each robot command depend on a generated future. motivation
2.2 What this reading follows
A robot can benefit from learning what its actions will look like without rendering those consequences before every movement. GigaWorld-Policy makes that distinction explicit: one video-derived Transformer learns both actions and their visual consequences, while an attention mask prevents the action stream from consulting the future stream. This permits action-only inference. The reported result is a favorable latency–score tradeoff against Motus on the tested PiPER tasks, plus stronger target-task data efficiency than pi0.5. Read the empirical claims with their conditions attached: physical scores use graded credit and repeated attempts, and several simulation and ablation numbers disagree within this revision. motivationarchitectureinferencelatencyreal-resultsevaluationdata-efficiencysimulationfuture-ablation
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | One Model |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
Shared Transformer blocks and Q/K/V weights directly support One Model; joint video/action training and optional future decoding support both recorded subcategories. Other mechanisms and Outside quadrants are reasonable catalog conventions for action-first forward dynamics without mandatory video generation or an inverse-dynamics stage. The paper does not define these quadrant labels. architecturefactorizationinferencemotivation
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 Make actions causally prior to their visual consequences
Begin with the two conditionals in Eqs. (2)–(3). The first produces an action chunk and conditioning latent c_t from current observations, proprioception and language. The second predicts visual consequences conditioned on that action signal. Figure 4 makes this ordering operational: an action query cannot read future-image keys, although future queries can read action keys. The same Transformer parameters serve both streams, so this is not a separate video planner followed by an inverse-dynamics controller. At deployment, Eq. (12) integrates noisy action tokens while the context contains only language, state and observations. After commands are executed, fresh sensor readings close the loop. Optional future-video generation can remain useful for inspection, but the specified action distribution does not require it as an input. factorizationarchitectureinference
Figure 4. Read rows as queries: actions cannot consult future-video keys. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the arrows to orient the matrix: Q runs downward through query rows, and K runs rightward through key columns. Green means attention is allowed; white means it is blocked. The first two rows allow observation and state to communicate while excluding actions and future images. The action row includes observation, state and action columns, but its future column is white. The future row is entirely green. The caption says actions attend to states and observations “only,” yet the graphic also colors the action-to-action diagonal green. Read that wording as excluding future conditioning, not as deleting within-action attention. Section 3.2 and Eqs. (2)–(3) support the directional dependency. architecturefactorizationobjectivesinference
What it supports. Future-image tokens can use action information, while action tokens do not need future-image information. This asymmetry explains why dropping the future stream can preserve the action model’s required inputs. Sharing weights still lets the visual objective affect learning; the mask is an information-flow constraint, not a stop-gradient operation.
Where the evidence stops. Language is absent from this matrix because it enters through cross-attention. The caption’s “only” wording is less explicit than the green self-attention diagonal; the faithful crop is retained rather than silently modifying the mask.
5.2 Why sparse futures can teach a policy that runs without them
The training objective helps explain why removing video at inference need not remove its training benefit. Pre-training fits visual dynamics; post-training combines video and action velocity losses with weights one and five. Because their Transformer parameters are shared, visual errors can change the representation subsequently used for actions. This is an interpretation of the architecture, not proof that it acquires a particular physical law. The default horizon contains 48 actions and four sampled future observations at stride 12. Table 5 favors that stride over denser or sparser alternatives. However, its no-video entry is 0.60 while the text says 0.65, and its zero-stride label cannot enter the frame-count formula. This supervision ablation should not be confused with simply disabling video decoding in an already trained policy. objectivesarchitectureinferencefuture-ablation
5.3 Separate prediction quality, command latency and robot success
Three kinds of evidence answer different questions. Table 6 measures both physical SR and future-image reconstruction, so better PSNR or SSIM is evidence about the image stream, not automatically about executed actions. Table 3 measures model-inference latency, which can affect feedback but omits a complete sensing-to-actuation timing breakdown. Table 4 measures four physical tasks under graded credit and repeated attempts. The authors attribute some advantage to faster feedback, yet comparing different policies changes learned priors and model design alongside delay. Reader analysis therefore treats latency as a plausible contributor requiring a controlled test. Similarly, Figure 7’s five-versus-fifty demonstration comparison concerns downstream task data; the preceding embodied pre-training remains part of the method’s cost. None of these distinctions negates the reported scores, but each changes what a reproduction must measure. mask-ablationlatencyevaluationreal-resultsdata-efficiencycurriculum
5.4 Training and inference
During training
Initialize from web-video-pretrained Wan, adapt with approximately 10,000 hours of embodied videos, then post-train on target-robot trajectories. Pre-training uses video flow matching only; post-training adds action flow matching with action/video weights 5/1. curriculumobjectivesinference
Video loss conditions on action tokens; action loss excludes future tokens. Shared parameters receive auxiliary visual supervision without future-to-action attention. The exact handling of missing action labels in human-video pre-training is unspecified. architectureobjectivescurriculum
During inference
Encode fresh observations, state and language; sample noisy actions and integrate their velocity field into continuous commands. Execute actions, acquire feedback and repeat. Reported latency excludes future-video tokens; the executed prefix before replanning is unspecified. inference
Optional future video can be jointly denoised with actions or generated afterward using the saved action-denoising KV cache. This optional prediction is distinct from physical execution. inference
5.5 Implementation flow
- Build a common representation
Compose the three views into one image; encode current and future images with the same pretrained VAE. Linear projections embed states and actions. Visual tokens use 2D positions and state/action tokens use 1D temporal positions. Language enters separately through cross-attention. architecture
- Constrain information flow
Wan 2.2 5B shares Transformer blocks and Q/K/V matrices. Observation/state tokens cannot read actions or futures. Actions read observations, states and other action tokens, but cannot read futures. Futures can read all four groups. Figure 4 retains within-action attention despite the caption’s abbreviated “only” wording. architectureinference
- Learn action-conditioned consequences
The same model predicts actions and an action latent signal, then models future observations conditioned on that signal. A 48-step chunk and stride 12 give four future observations. No inverse-dynamics decoder or candidate-rollout action search is specified. factorizationinference
6. Experiments & results
GigaWorld-Policy turns a video diffusion Transformer into a robot policy that learns actions alongside action-conditioned future visuals, then omits future-video generation during control. A shared causal mask makes this separation architectural. Physical experiments favor its speed–score tradeoff, although graded evaluation and inconsistent source numbers require care.
6.1 Read the original evidence
Table 3. The action-only model is much faster than Motus, while pi0.5 remains faster. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the Time column: it reports milliseconds on an NVIDIA A100, so smaller values indicate faster model inference. The method’s 360 ms measurement uses action-only decoding. Next inspect the two SR columns separately: simulation uses binary task completion, whereas the physical protocol allows graded credit. Blank simulation entries for GigaBrain-0 and Cosmos-Policy are missing results, not zeros. The Ours and Motus rows support the approximately ninefold latency comparison. The pi0.5 row provides a different tradeoff, with lower latency but lower reported scores. Do not replace this table’s simulation entries with the clean/randomized averages from Tables 2 or 8. latencyinferenceevaluationsimulation
What it supports. Dividing 3231 ms by 360 ms gives approximately 9× faster inference than Motus, alongside reported physical scores of 0.83 versus 0.76. Against pi0.5, the method takes 1.6× as long: 360 versus 225 ms. The table therefore establishes a particular speed–score tradeoff, not universal speed dominance.
Where the evidence stops. These are model-inference timings, not a measured complete sensing-to-actuation cycle. Real-world SR is graded and permits repeated attempts. Table 3’s pi0.5 simulation value also differs from the separate clean/randomized summaries, and no reconciliation is supplied.
Table 4. Physical task scores favor GigaWorld-Policy under the paper’s graded evaluation protocol. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across the final row before looking at the average. The four columns describe different manipulations: placing tableware in a basket, aligning an object with a QR scanner, sweeping debris and nesting bowls. Appendix C provides the task constraints, including plates beneath bowls during cleaning and returning the scanned object to its original location. Ours scores 0.90, 0.75, 0.75 and 0.90 respectively. Compare each value with the Motus row: QR scanning ties, while the other three entries improve. The final column rounds the average to two decimals. Do not confuse the fifty training demonstrations per task with the twenty evaluation trials. real-resultsevaluationtasks
What it supports. The table reports average SR 0.83 versus Motus at 0.76 and pi0.5 at 0.69. Those rounded averages differ by seven and fourteen percentage points. The individual columns show that the advantage is not uniform: the QR-scanning comparison with Motus is a tie.
Where the evidence stops. Each method receives 20 trials per task and up to five attempts per trial. The paper illustrates graded scoring through grasp/place credit but omits a full task-specific rubric, so these numbers cannot be read as first-attempt binary completion frequencies.
Figure 7. Five task demonstrations reach the baseline’s fifty-demonstration score. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the legend to separate the yellow pi0.5 bars from the green GigaWorld-Policy bars. The horizontal axis is the number of task demonstrations, with three tested budgets: five, twenty-five and fifty. The labels above each bar give the reported score, avoiding the need to estimate heights. For the paper’s 10% claim, compare across groups: the green bar at five demonstrations and the yellow bar at fifty are both 69%. Comparing bars within a group instead asks which model learns more from an equal target-task budget. Appendix C establishes fifty trajectories per full real-world task dataset. data-efficiencytasksevaluationcurriculum
What it supports. GigaWorld-Policy reports 69%, 78% and 83% across increasing task-data budgets, compared with 46%, 58% and 69% for pi0.5. The observed five-versus-fifty match supports improved target-task data efficiency in this setup, while the equal-budget comparisons favor the proposed model at every plotted point.
Where the evidence stops. The smaller budget does not include the method’s web-video initialization and embodied-video pre-training. No error bars or repeated-run variation are shown. Matching one plotted score does not establish a general tenfold reduction in total data or training cost.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Action-only inference latency on NVIDIA A100 Per-step model inference; future-video decoding disabled. | 360 ms Latency (ms), lower is better | Motus: 3231 ms; pi0.5: 225 ms. About 9× faster than Motus, but 1.6× slower than pi0.5. These are inference timings, not measured end-to-end robot cycle times. latencyinference |
| Four real-world PiPER manipulation tasks 50 post-training demonstrations per task; 20 trials per task, up to five attempts per trial; graded real-world scoring. | Average 0.83; cleaning 0.90, QR scanning 0.75, sweeping 0.75, bowl stacking 0.90. Reported SR, higher is better | Motus average 0.76; pi0.5 0.69. Rounded averages differ by 7 and 14 percentage points. QR scanning ties Motus. This is not one-attempt binary success. evaluationreal-resultstasks |
| RoboTwin 2.0 clean and randomized evaluation 50 tasks; 2,500 clean plus 25,000 randomized training demonstrations; 100 test episodes per task. | Table 8: 0.86 clean / 0.85 randomized. Table 2 instead reports 0.87 / 0.85. Binary task-success rate | Motus 0.89 / 0.87; X-VLA 0.73 / 0.73; pi0.5 0.43 / 0.44. Above listed VLAs but below Motus. The clean-score discrepancy remains unresolved; Table 3 separately gives pi0.5 simulation SR as 0.48. simulationevaluationlatency |
| Real-world demonstration efficiency Per-task post-training subsets of 5, 25 and 50 demonstrations; same stated evaluation protocol. | GigaWorld-Policy: 0.69, 0.78, 0.83. Reported SR | pi0.5: 0.46, 0.58, 0.69. Five demonstrations match the baseline’s fifty-demonstration score. The tenfold reduction concerns target-task data, not total training data or compute. data-efficiencyevaluationcurriculum |
| Causal-mask ablation Unconstrained self-attention comparison; physical SR and held-out target-platform video reconstruction. | 0.83 / 28.41 / 0.901 SR / PSNR / SSIM, higher is better | Self-Attn: 0.81 / 27.87 / 0.892. A small SR difference accompanies improved reconstruction. No uncertainty is reported; reconstruction alone does not establish better executed control. mask-ablation |
| Pre-training and future-supervision ablations Physical-task ablations; initialization and embodied pre-training, then varied future stride with a fixed 48-step action horizon. | Both pre-training stages: 0.83; stride 12: 0.83. Reported SR | Scratch/video-only/embodied-only initialization: 0.45/0.57/0.73. Table 5 no-video column: 0.60; accompanying text: 0.65. The reported optimum uses moderate future sampling. Preserve the conflicting baseline and treat the zero-stride label as a special condition, not division by zero. pretraining-ablationfuture-ablation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 6. The causal mask changes deployment dependencies while retaining similar reported task performance. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the row labels: Self-Attn is the unconstrained variant in which all token types may attend to one another; Ours uses the causal mask from Figure 4. Read SR separately from PSNR and SSIM. SR summarizes physical-task evaluation, while the other two columns assess reconstructed future images on a held-out target-platform test set. All three arrows point upward. The numbers favor the causal version, but the action score moves only from 0.81 to 0.83. Figure 9 supplies selected visual examples of predicted object details; it is illustrative support rather than another aggregate performance measure. mask-ablationarchitectureinferenceevaluation
What it supports. Ours reports 0.83 SR, 28.41 PSNR and 0.901 SSIM, versus 0.81, 27.87 and 0.892. The central architectural benefit is that future-video prediction can be omitted at inference. The small score difference accompanies that benefit but does not by itself prove a statistically reliable control improvement.
Where the evidence stops. The table supplies neither uncertainty nor a latency column for these two ablation variants. Higher reconstruction scores do not isolate why physical control changes, and selected red-box examples in Figure 9 cannot establish a general causal link between image fidelity and task success.
7. Analysis & limitations
7.1 What the evidence leaves open
The paper does not isolate latency as the cause of better physical scores: models, priors and decoding differ together. Graded credit, repeated attempts and missing uncertainty limit comparisons with conventional binary success rates. evaluationlatencyreal-results
Repeated simulation cells also conflict: Motus Place Container Plate randomized is 0.99 in Table 2 versus 0.96 in Table 8; Stack Blocks Two clean is 1.00 versus 0.99. These and headline gains should not be silently reconciled. simulationmotivation
Evidence covers four physical tasks on one platform and trained simulation tasks, not broad unseen-embodiment transfer. Table 8’s Hanging Mug scores of 0.16/0.12 show task-specific weakness despite high averages. simulationreal-resultstasks
7.2 Questions for discussion
- Would equalizing observation-to-command delay preserve the physical-score advantage?
- How sensitive is action-only control to the future-supervision weight and embodied pre-training coverage?
8. Reproducibility audit
8.1 Requirements and known gaps
Appendix A specifies 6000 GPU hours, batch size 256, AdamW beta values 0.85/0.9 and cosine learning-rate decay from 1e-4 to 1e-6. A100 is identified for latency measurement, not as training hardware. implementationlatency
Reproduction still needs cleaning/sampling rules, action normalization, image resolution, frozen-module choices, post-training schedule, denoising solver/steps, training GPU type/count and executed action prefix. Complete task-specific grading and trial-level uncertainty are absent. curriculumarchitectureinferenceimplementationevaluationtasks
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 deleting future tokens leave the action computation unchanged?
Reader-proposed check, not performed: hold one causal-model checkpoint, observation/state/language context, initial action noise, solver schedule and numerical precision fixed. Compare action-only decoding with joint decoding using two different future-noise samples. Record action velocity outputs at every solver step and final actions, with a numerical tolerance fixed in advance. The causal dependency predicts agreement within that tolerance; systematic sensitivity to future noise would reveal an unintended input path. Use an unconstrained-attention checkpoint as a comparison where such invariance is not architecturally guaranteed. Measure latency separately from numerical agreement. architectureinferencemask-ablation
Check 2: How much of the physical-score difference comes from feedback delay?
Reader-proposed check, not performed: evaluate the same GigaWorld-Policy checkpoint at its native timing and with added delay targeting Motus’s reported 3231 ms, holding executed chunk prefix, sensor pipeline and task instructions fixed. Counterbalance trial order and pair initial object layouts across conditions on the four PiPER tasks. Record actual observation-to-command delay, strict first-attempt completion and the paper’s graded/repeated-attempt score separately. Compare with Motus under a matched-delay protocol if its implementation is available. A substantial within-policy loss under added delay supports the feedback explanation; little change would weaken latency as the main explanation. Report trial-level uncertainty. latencyinferenceevaluationreal-resultstasks
8.3 Reading coverage
Visual audit: The title and complete 24-name byline were inspected on page 1. Visual reading covered Figures 1–10 and Tables 1–8 on the declared pages, including all equations, training/inference details, evaluation conditions and Appendix A compute settings cited here. Pages 5 and 11 resolved extraction-damaged labels. Figure 4’s query/key orientation and within-action diagonal were checked against its caption and Section 3.2; the caption’s abbreviated wording is disclosed. Tables 2 and 8, and Table 5 versus its discussion, were compared without reconciling conflicting values. Each of the six final original crops was opened and visually inspected with its labels, legends or table headers intact. Page 3 and reference pages 16–20 were read as text but were not visually inspected. No external supplement or code was inspected, and no reproduction check was run.
PDF pages inspected for this edition: 1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15. 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
- 1. Introduction
- 2. Related Work, Sections 2.1–2.2
- 3. Method, Sections 3.1–3.4 and Eqs. (1)–(12)
- 4. Experiment, metrics, implementation and Sections 4.1–4.5
- 5. Conclusion
- Supplementary material: Appendices A–C and Table 8
- References, PDF pp. 16–20
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.
- Some PDF font mappings yielded unpaired Unicode surrogates; replacement characters are recorded by page in extractionNotes and require PDF visual verification.
- All six supplied text chunks were read. Original PDF pages 1, 2 and 4–15 were visually inspected; this resolves the retained Figure 4 and Figure 7 labels affected by extraction on pages 5 and 11. The embedded supplementary material was read; no separately hosted supplement was inspected.
- Identity/version scope: the title and 24-person GigaWorld Team roster match the catalog. The inspected artifact is arXiv:2603.17240v2, dated 21 March 2026 in the margin, with a 24 March masthead. The catalog submission date is 18 March. These are distinct date fields; v1 was not supplied or compared. The manifest identifies earlier HTML as the same v2, but that HTML was not read in this bundle.
- Code, project-page resources and external datasets were not inspected; no experiments were reproduced. The source has no dedicated limitations section.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title, GigaWorld Team byline, masthead and arXiv margin
The exact catalog title and all 24 listed team members match the PDF. The margin identifies arXiv:2603.17240v2, 21 March 2026; the masthead reads 2026-3-24. GigaAI is printed beneath the title.
Go to primary source ↓motivationPDF pp. 1–2, Introduction and Figure 2
The authors contrast sparse action supervision with dense visual supervision and distinguish their optional future-video branch from mandatory joint generation and video-to-inverse-dynamics pipelines.
Go to primary source ↓factorizationPDF p. 4, Figure 3 and Section 3.1, Eqs. (2)–(3)
One model generates an action chunk and action latent conditioning signal from observations, proprioception and language; future observations are conditioned on that signal. The future stride is Delta, with K=floor(p/Delta). Figure 3 separates joint training from action inference and optional video decoding.
Go to primary source ↓architecturePDF p. 5, Section 3.2, Eqs. (4)–(5), Figure 4; p. 6, opening paragraph
Three views are composed into one image. A shared VAE encodes current/future images; states/actions use linear projections. All modalities share Transformer blocks and Q/K/V projections. Figure 4 has green action-to-action and future-to-future diagonal blocks but a white action-query/future-key block. Language enters through cross-attention, outside the self-attention sequence.
Go to primary source ↓objectivesPDF pp. 6–7, Section 3.3, Eqs. (6)–(10)
Flow matching interpolates Gaussian noise and data and predicts their difference as velocity. Video prediction conditions on state, observation, action and language; action prediction conditions on state, observation and language. Pre-training uses video loss only; post-training combines weighted video and action losses.
Go to primary source ↓curriculumPDF p. 6, Table 1 and Section 3.3, Training Process and Data
Web-video model initialization is followed by embodied-video pre-training on approximately 10,000 hours, then target-robot image/language/action trajectories. Table 1 names EgoDex, Agibot, EGO4D, RoboMind, RDT, Open X-Embodiment, DROID, ATARA and Something-Something V2. Cleaning, formatting and sampling are mentioned without a detailed recipe.
Go to primary source ↓inferencePDF p. 7, Section 3.4, Eqs. (11)–(12), and Implementation Details
Action-only inference samples Gaussian action tokens and integrates the learned velocity field with language/state/observation context, then executes actions and refreshes observations. Optional video uses joint denoising or the saved action-denoising KV cache. The backbone is Wan 2.2 5B; p=48, Delta=12, lambda_action=5 and lambda_video=1. Reported latency disables future-video decoding.
Go to primary source ↓evaluationPDF p. 7, Evaluation Metrics and Implementation Details; p. 8, continuation; p. 9, Section 4.2; p. 14, Appendix C
Simulation success is binary, with 100 test episodes per task. Real-world evaluation uses graded credit, illustrated by 0.5 for grasp and 0.5 for placement, with 20 trials per task and up to five attempts per trial. Simulation training mixes 2,500 clean and 25,000 randomized demonstrations across 50 tasks. Real-world post-training uses 50 trajectories per task.
Go to primary source ↓latencyPDF p. 8, Table 3 and Section 4.1
On NVIDIA A100, Ours reports 360 ms, 0.86 simulation SR and 0.83 real-world SR; Motus 3231 ms, 0.88 and 0.76; pi0.5 225 ms, 0.48 and 0.69. GigaBrain-0 reports 452 ms and 0.68 real-world SR, and Cosmos-Policy 1413 ms and 0.58, with simulation values absent.
Go to primary source ↓simulationPDF p. 8, Table 2; p. 9, Section 4.2; p. 15, Table 8
Table 2 averages are Ours 0.87/0.85, Motus 0.89/0.87, X-VLA 0.73/0.73 and pi0.5 0.43/0.44 for clean/randomized scenes. Table 8 instead gives Ours 0.86/0.85. Repeated Motus cells also differ: Place Container Plate randomized 0.99 versus 0.96, and Stack Blocks Two clean 1.00 versus 0.99. Table 8 shows Ours Hanging Mug 0.16/0.12. Section 4.2 names background, clutter, table-height and lighting randomization.
Go to primary source ↓real-resultsPDF p. 9, Table 4 and Section 4.3
AgileX PiPER 6-DoF manipulation results for Ours are Clean the Desk 0.90, Scan a QR Code 0.75, Sweep up Trash 0.75 and Stack Bowls 0.90, with rounded average 0.83. Motus reports 0.80/0.75/0.70/0.80, average 0.76; pi0.5 averages 0.69. The QR task ties Motus.
Go to primary source ↓data-efficiencyPDF p. 10, Section 4.4; p. 11, Figure 7 and caption
Figure 7 plots demonstration counts 5, 25 and 50. GigaWorld-Policy scores 69%, 78% and 83%; pi0.5 scores 46%, 58% and 69%. The 10% claim compares five task demonstrations with fifty, not total pre-training data.
Go to primary source ↓future-ablationPDF p. 10, Table 5 and caption; p. 11, Impact of the Number of Predicted Future Frames
For the table columns labeled Delta=0,4,8,12,24,48, SR is 0.60,0.76,0.78,0.83,0.80,0.76. The caption separately defines K=0 as no video prediction; Delta=0 cannot be substituted into floor(48/Delta). The text instead states improvement from 0.65 to 0.83, an unresolved baseline discrepancy.
Go to primary source ↓mask-ablationPDF p. 11, Table 6 and Role of Causal Self-Attention; p. 12, Figure 9 and continuation
Unconstrained Self-Attn scores SR 0.81, PSNR 27.87 and SSIM 0.892; Ours scores 0.83, 28.41 and 0.901 on the real-world evaluation/test set. The text emphasizes optional video inference and attributes improved dynamics to the mask. Figure 9 shows selected ground-truth, Self-Attn and Ours images with red detail boxes.
Go to primary source ↓pretraining-ablationPDF pp. 10–11, Importance of Pre-training, Table 7 and Figure 8
Table 7 reports SR 0.45 from scratch, 0.57 with video initialization only, 0.73 with embodied pre-training only and 0.83 with both. Figure 8 increases from 57% at zero embodied data to 83% at the full fraction, holding video initialization and post-training settings fixed.
Go to primary source ↓implementationPDF p. 13, Appendix A, Pretraining Details
The appendix specifies 6000 GPU hours, global batch size 256, AdamW beta1=0.85 and beta2=0.9, and cosine learning-rate decay from 1e-4 to 1e-6. It does not identify training GPU type/count, denoising steps or complete post-training configuration. The appendix also contains an unexplained SwiftWA name and an unresolved question-mark citation after AdamW.
Go to primary source ↓tasksPDF p. 14, Appendix C and Figure 10; p. 9, Figure 5; p. 10, Figure 6
The four physical tasks use 50 demonstrations each. Desk cleaning requires plates below bowls in a basket; bowl stacking requires nesting; QR scanning requires tool pickup, object pickup, alignment and return; sweeping requires coordinated brush/dustpan manipulation. Figures show selected successful deployment sequences, not trial-level statistics.
Go to primary source ↓8.5 Primary sources
GigaWorld-Policy: An Efficient Action-Centered World–Action Model ↗
PDF · 9,781 extracted words
Source fingerprint
786a3e27a8bbe7e97f3a29020f78681d2a11a320f0a717dbe66930f7635c8ad4