Foresight Without Seeing: Latent Futures for World Action Models
1. Paper overview
In one sentence: ForeWAM reuses a single video-feature prefill to condition direct action denoising, achieving efficient benchmark control while leaving the causal meaning and broader robustness of its latent futures unresolved. e2e3e4e5e6e9e10e11e12
| At a glance | What to know |
|---|---|
| Research problem | Source description Generating future video can expose scene evolution to a policy but adds iterative inference cost. ForeWAM asks whether a direct action policy can access predictive dynamics through hidden features while dispensing with future-video generation. Its intended contribution is the combination of a distributed cache and compact, transition-supervised registers. e2e3 |
| Core mechanism | |
| A key reported result | Observed LIBERO-Plus robustness: ForeWAM 61.6; Flash 58.2. Observed task success rate (%). Observed subset spanning seven perturbation categories; external baseline results come from different sources. Fast-WAM 51.5: differences are +10.1 and +6.7 percentage points. pi0.5 reports 85.7 overall. ForeWAM exceeds Fast-WAM in camera (62.5 versus 16.4) and noise (58.8 versus 37.7), but trails it in robot-initial-state shifts (37.4 versus 44.5) and lighting (74.1 versus 78.2). Gains are descriptive and not universal. e8e10 |
| Reading caution | Source description All evaluated control is confined to LIBERO and LIBERO-Plus. Real-world transfer, other robot morphologies and broader long-horizon dynamics remain unverified. e13 |
Core contributions
Figure 2. A video backbone and an action expert share predictive context; the left teacher branch supplies training supervision. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the bottom: the VAE encodes observations, the action encoder supplies action tokens, and the text encoder conditions the central model. The blue Video DiT and green Action DiT remain distinct branches. Follow the purple dynamics token through the adapter toward the latent-action target at left. The snowflake marks the frozen encoder path. The left-pointing Distill arrow illustrates target matching; Eq. (9) specifies the actual learning boundary, with stop-gradient on the teacher target. The caption separates deployment from this training overview: future slots contain noise, the teacher path disappears, and a single video prefill supplies cached K/V for action denoising. e3e5e6e7
What it supports. The architectural contribution is a way to condition the action expert without producing a future video at deployment. Distributed video K/V and compact dynamics registers both reach the policy. The inverse-dynamics branch provides a training target; executable motor commands emerge from the Action DiT.
Where the evidence stops. Do not read the illustrated future frames or IDM block as deployment inputs. Figure 2 is a training overview, and Eqs. (3)–(5) and (9) are needed to distinguish stochastic future slots, fixed teacher targets and action outputs.
2. Motivation
2.1 The problem and the proposed response
Generating future video can expose scene evolution to a policy but adds iterative inference cost. ForeWAM asks whether a direct action policy can access predictive dynamics through hidden features while dispensing with future-video generation. Its intended contribution is the combination of a distributed cache and compact, transition-supervised registers. e2e3
2.2 What this reading follows
A robot may benefit from predicting how an interaction unfolds, yet generating a video before every action can make control expensive. ForeWAM puts that predictive context inside the policy's hidden computation. It keeps the current visual latent, appends stochastic future slots, and caches one pass through a Video DiT for use by an Action DiT. A frozen latent-action teacher shapes compact dynamics registers during training and disappears at deployment. The six visuals below explain that information flow, the mask that makes cache reuse possible, and the distinction between standard-suite success, observed-subset robustness, standalone latency and component evidence. e2e3e4e5e6e9e10e11e12
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 | IDM |
| Quadrant | Q4 · Dual-system × IDM |
3.1 Evidence-based assessment
Conflicts with the recorded classification
The separate Video DiT and Action DiT support Dual-system. However, IDM describes the frozen training teacher, not deployed action extraction: ForeWAM directly denoises executable actions from cached features. Therefore Q4/Dual-system × IDM is misleading as an inference-paradigm label. Joint training also does not make this a single shared model or joint future/action generation at deployment. e3e4e5e6e14
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 Build predictive context without observing a future
Begin with the deployment inputs in Eq. (1): observation o, instruction l and proprioception p. The VAE supplies a clean current-frame latent. Eq. (3) appends Gaussian noise in the positions that would otherwise hold future visual information. A single Video DiT prefill at sigma = 1.0 creates register states and layer-wise K/V tensors; the Action DiT reads them throughout action denoising. Training makes this interface meaningful by supplying demonstrated future latents to the video flow objective, while keeping action-loss gradients connected to the video-to-action path. Deployment supplies neither demonstrated future frames nor teacher outputs. Thus a latent future here is an internal conditioning state, not a generated observation. The Action DiT outputs the executable chunk directly, with ten denoising steps in ForeWAM and two in Flash. e3e5e7e11
Figure 3. The mask gives action tokens broad access while keeping the video prefill independent of action-token updates. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read rows as querying groups and columns as the groups they can access, consistent with the paragraph below Figure 3. The current frame f0 corresponds to C; D denotes dynamics registers; f1 through fn stand for future tokens; a1 through an denote actions. Filled blue cells show restricted video-side routing: current tokens see current tokens, registers see current/register tokens, and later future tokens can read earlier video context. The green action rows span every column. Each denoising step combines fresh action K/V with cached video K/V. Crucially, the D row has empty future columns: the registers are not pooling observed future tokens through this mask. e4e5e6e12
What it supports. This asymmetric routing makes a reusable video context compatible with repeatedly updated action tokens. It also distinguishes the two interfaces: future-slot features provide distributed context, whereas registers learn transition-oriented cues from their permitted context under teacher supervision. Their roles should not be collapsed into a single future-observation input.
Where the evidence stops. Filled cells specify permitted attention, not measured causal influence or accurate physical dynamics. The caption and text support the routing interpretation, but neither this figure nor the aggregate ablation identifies which tokens actually determine successful actions.
5.2 Separate what the registers can read from what they learn
Figure 3 makes an easily missed distinction: dynamics-register queries can access current tokens and registers, but their future-slot columns are blank. Future tokens, in turn, can access the registers, and action tokens can access every group. Eq. (9) explains how transition supervision reaches the registers without giving them observed futures at deployment. A frozen teacher encodes a demonstrated transition into a non-executable latent target; a projection of the mean register state is trained to match it. Stop-gradient applies to that target, not the action-to-video learning path. A reader interpretation is that the registers learn to predict transition-oriented cues from allowed context, while future slots offer a distributed complementary interface. This interpretation follows the routing and loss; it is not proof that particular registers encode contact or that the action expert needs them. e4e5e6e7
5.3 Read the evidence as several different comparisons
Start with Table 1: ForeWAM and Flash achieve similar standard-LIBERO averages, 96.7% and 96.9%. Move to Table 2 and the same variants separate, at 61.6% and 58.2% on the observed LIBERO-Plus subset. Table 3 then provides a possible reason to accept that tradeoff: standalone latency drops from 568 ms to 220 ms on the specified A800. These metrics address different questions and should not be fused into a single superiority claim. Finally, Table 4 compares the mechanisms: both components exceed either alone among three coverage-matched configurations. Its base row has different coverage, so it does not complete a controlled factorial ablation. My assessment is that the evidence supports a useful configuration and a speed–robustness tradeoff, while leaving causal feature semantics and broader deployment generalization open. e9e10e11e12e13
5.4 Training and inference
During training
Initialize the visual branch from Wan2.1-T2V-1.3B and the action expert from a linearly interpolated Wan2.1 ActionDiT checkpoint. Both have 30 blocks, with widths 1536 and 1024. Use 16 dynamics registers and a 32D teacher target. No reported variant receives embodied pretraining before LIBERO training; the video initialization is still pretrained. e14e15
Each example contains 33 observation frames; temporal ratio four maps the 32-step action chunk to nine video frames. Joint training combines future-latent flow, action flow and register matching. AdamW uses learning rate 1e-4, weight decay 0.01, cosine annealing and clipping at 1.0; the flow schedule has 1,000 timesteps and shift 5.0. e7e14e15
During inference
From the current observation, instruction and proprioception, prefill at sigma = 1.0 and reuse that cache for ten action-denoising steps. Flash uses OneDP distillation to reduce the action schedule to two steps while retaining the interface. Neither observed future frames nor teacher outputs are available. e2e3e5e11
The output is a control chunk, not a video plan requiring an inverse-dynamics decoder. The paper describes chunk-level queries but leaves the executed prefix length and exact observation-refresh cadence unspecified. e3e5e14
5.5 Implementation flow
- Encode observation and instruction
Concatenate two 224×224 camera views into a 224×448 image before VAE encoding. Use Wan2.1 text embeddings and project proprioception into a context token. Video and action branches have separate transformer stacks. e4e14
- Build hidden predictive context
Keep the current visual latent clean, append Gaussian future slots, and prefill the Video DiT once. Cache layer-wise keys and values and retain dynamics-register states. These hidden features are not decoded future observations. e3e5
- Route information into action denoising
Figure 3 shows current tokens reading current tokens, registers reading current/register tokens, and future tokens using causal temporal access. Action queries read all video/register groups and action tokens. In particular, registers do not read future slots under the depicted mask. e4
- Shape the compact interface
During training, a frozen inverse-dynamics teacher encodes a demonstrated visual transition. Mean-pool the dynamics registers and project them toward its latent target. The teacher supplies a non-executable learning target; the Action DiT supplies motor commands. e6
6. Experiments & results
ForeWAM gives a direct robot policy hidden predictive context through one Video DiT prefill, then reuses its K/V cache while an Action DiT denoises executable actions. A frozen teacher shapes dynamics registers only during training. Reported success is 96.7% on standard LIBERO and 61.6% on the observed LIBERO-Plus subset; Flash trades some subset robustness for lower latency. These are benchmark control results, with explicit coverage limits (e3, e5, e6, e9–e11).
6.1 Read the original evidence
Table 1. The accelerated variant retains standard-suite success in the reported benchmark evaluation. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First read Params and Embodied PT to distinguish parameter count from prior robot-data training. The two highlighted ForeWAM rows both list 2B parameters and no embodied pretraining. Then compare Spatial, Object, Goal and Long before looking at Overall. The paper evaluates each task with 50 rollouts, and all success entries are percentages. ForeWAM's Long score, 92.8, is lower than its Object score, 99.6; the overall average hides that spread. Flash reaches 96.9 overall versus 96.7 for the standard variant. Fast-WAM remains higher at 97.6 overall, so efficiency and standard-suite performance must be considered together. e8e9e11e14e15
What it supports. Reducing the action-denoising schedule does not produce a large aggregate standard-LIBERO drop in these reported runs. Flash is 0.2 percentage points above ForeWAM overall, while both trail Fast-WAM. The 2B-versus-6B parameter contrast provides context, but does not isolate the effect of model size.
Where the evidence stops. No uncertainty intervals appear in the table, so the small difference between ForeWAM variants is not evidence of statistical superiority. No embodied pretraining also does not mean training from scratch: Appendix A specifies pretrained Wan components.
Table 2. Observed robustness gains vary sharply by perturbation category and comparison method. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat Camera, Robot, Language, Light, Background, Noise and Layout as distinct perturbation categories, with percentage success increasing upward. Ours is ForeWAM and Ours-Flash is its accelerated variant. Compare category columns before Overall: ForeWAM improves camera success from Fast-WAM's 16.4 to 62.5 and noise success from 37.7 to 58.8. However, its robot-initial-state and lighting entries are lower than Fast-WAM's. The overall 61.6 therefore summarizes uneven behavior, not an across-category dominance result. The table also includes stronger overall results, notably pi0.5 at 85.7. Read it alongside Section 4.2's explicit warning about external sources and subset coverage. e8e9e10e13
What it supports. Relative to the reported Fast-WAM result, ForeWAM improves observed overall success by 10.1 percentage points; Flash improves it by 6.7. Flash is 3.4 points below the standard variant on this subset, exposing a robustness cost that the nearly unchanged standard-LIBERO average does not reveal.
Where the evidence stops. These are observed-subset comparisons, and external baseline results come from different sources. The paper does not provide a common evaluation-ID list that makes all rows coverage-matched; the table cannot establish universal robustness or a causal architecture advantage.
Table 3. Two-step action denoising lowers measured standalone latency while preserving the Future-KV interface. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The header gives milliseconds and a downward arrow: smaller values indicate less time to generate an action chunk. Section 4.3 supplies the measurement context omitted from the compact table: one NVIDIA A800 GPU with 80 GB memory. ForeWAM's standard action schedule uses ten denoising steps, whereas Flash uses two. Both retain the one-prefill cache interface, so Flash's additional acceleration concerns the action branch. Compare 568 ms with 220 ms to evaluate that schedule change, and compare with Fast-WAM's 667 ms for the reported external configuration. Keep this standalone measurement separate from benchmark success and elapsed task-completion time. e5e9e10e11
What it supports. Flash reduces reported latency by about 61% relative to ForeWAM. The practical tradeoff is visible only by combining this table with success results: standard LIBERO remains similar, while the observed LIBERO-Plus subset falls from 61.6% to 58.2%. Fast inference alone does not establish stronger control.
Where the evidence stops. The table does not report latency distributions or a full control-loop timing breakdown. It cannot establish achieved robot control frequency, end-to-end task speed or behavior on other hardware; the A800 specification applies to inference rather than training compute.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Standard LIBERO manipulation Spatial, Object, Goal and Long; 50 rollouts per task; no embodied pretraining for ForeWAM variants. | ForeWAM 96.7 overall; Flash 96.9. ForeWAM suite rates: 97.0, 99.6, 97.2, 92.8. Task success rate (%) | Fast-WAM 97.6 overall; ForeWAM and Flash are 0.9 and 0.7 percentage points lower. Flash preserves aggregate standard-suite performance in the reported runs. The table reports 2B parameters for either variant versus 6B for Fast-WAM; this is not a controlled size comparison. e8e9 |
| Observed LIBERO-Plus robustness Observed subset spanning seven perturbation categories; external baseline results come from different sources. | ForeWAM 61.6; Flash 58.2. Observed task success rate (%) | Fast-WAM 51.5: differences are +10.1 and +6.7 percentage points. pi0.5 reports 85.7 overall. ForeWAM exceeds Fast-WAM in camera (62.5 versus 16.4) and noise (58.8 versus 37.7), but trails it in robot-initial-state shifts (37.4 versus 44.5) and lighting (74.1 versus 78.2). Gains are descriptive and not universal. e8e10 |
| Standalone action-generation latency One NVIDIA A800 GPU with 80 GB memory; ten action steps for ForeWAM, two for Flash. | ForeWAM 568; Flash 220. Inference latency (ms; lower is better) | Fast-WAM 667; Flash has about 61% lower latency than ForeWAM. The measurement concerns action generation, not task-completion time, achieved control frequency or physical-robot throughput. e11 |
| LIBERO-Plus component comparison Three component configurations each have 1,482 coverage-matched observed evaluations; the base has 10,027 with different coverage. | Both components 61.6; Future-KV only 58.5; LA supervision only 58.0. Observed task success rate (%) | Combined configuration gains 3.1 and 3.6 percentage points over the single-component configurations. Base policy 53.6 is contextual. This supports configuration-level complementarity. A fully matched two-by-two experiment and causal interventions are still needed; subtracting the base row would confound coverage. e12 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. The combined interface leads the matched component configurations, while the base row has different evaluation coverage. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the four rows as configurations, with higher observed success preferred. Future-KV only retains the cache component without latent-action supervision; LA supervision only retains that supervision without Future-KV; Ours combines both. The highlighted combined score is 61.6%, compared with 58.5% and 58.0% for the two single-component configurations. Before subtracting the Base policy score of 53.6%, read the paragraph above the table. It states that the three component configurations each have 1,482 coverage-matched observed evaluations, whereas the base has 10,027 evaluations with a different coverage profile. Those counts are essential context that the compact crop itself cannot convey. e12e6e7
What it supports. The supported comparison is a 3.1-point advantage over Future-KV only and a 3.6-point advantage over LA supervision only within the matched configurations. This supports the usefulness of their combination at the configuration level. The base row is contextual evidence, not a matched estimate of the gain from adding both components.
Where the evidence stops. The table does not complete a matched two-by-two factorial experiment, supply uncertainty or intervene on learned states. It therefore cannot establish causal necessity, prove register semantics or attribute all improvement to a specific dynamics representation.
7. Analysis & limitations
7.1 What the evidence leaves open
All evaluated control is confined to LIBERO and LIBERO-Plus. Real-world transfer, other robot morphologies and broader long-horizon dynamics remain unverified. e13
Neither the mask nor aggregate success establishes that hidden future features represent accurate dynamics. Tables provide no uncertainty intervals, and external robustness baselines are not coverage-matched causal controls. e4e7e9e10e12
7.2 Questions for discussion
- Would the register advantage persist with identical subset coverage and several training seeds (e12)?
- Does intervention on future-slot K/V impair transition-sensitive tasks beyond the effect of removing generic context (e4–e7)?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the specified Wan components, action initialization, frozen latent-action teacher and synchronized LIBERO trajectories. Main text calls the teacher LaWM while Appendix A says LaWAM; the exact checkpoint and action-initialization procedure need clarification. e6e14e15
Missing details include the latent-action loss weight, batch size, training steps, training hardware/time, software versions, exact LIBERO-Plus evaluation IDs/seeds and Flash distillation settings. The A800 specification describes inference only. These gaps prevent treating the appendix as a complete executable recipe. e7e10e11e12e14e15
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: Complete the matched two-by-two component experiment
Reader-proposed check, not an experiment performed here: train base, Future-KV only, LA supervision only and both with identical Wan initialization, LIBERO demonstrations, optimization budget and ten-step action inference. Evaluate every checkpoint on exactly the same LIBERO-Plus task/perturbation/seed list and repeat training across several seeds. Report per-category success and uncertainty on paired configuration differences, including the interaction between the two components. The paper's configuration-level complementarity would be strengthened if both reliably exceed each single-component model under these controls. If the differences vanish or depend on coverage, the aggregate Table 4 ranking would not support a robust component advantage. e12e14e15
Check 2: Audit the deployment interface and the cost of cache reuse
Reader-proposed check, not a reproduction result: with a fixed observation, instruction, proprioception, future-slot noise and action noise, trace one action query. Verify that exactly one Video DiT prefill occurs, that its K/V is reused across action steps, and that no future frames or teacher outputs enter the deployment graph. Compare cached execution with recomputing the identical prefill from the same fixed inputs at every action step; action outputs should agree within numerical tolerance, while recomputation should add cost. Then compare ten-step ForeWAM and two-step Flash on the same held-out rollout seeds and A800 timing setup. Report prefill and denoising time separately alongside success, so a latency gain cannot conceal a changed input contract or robustness drop. e3e5e6e10e11e14
8.3 Reading coverage
Visual audit: Inspected the title, authors, affiliations and v1 date on page 1; Figure 1 and introductory method/OneDP claims on pages 2–3; Figure 2 and Eqs. (1)–(5) on pages 4–5; Figure 3, routing, cache prefill, teacher supervision and flow equations on page 6; Eqs. (9)–(10), Table 1 and evaluation/hardware context on page 7; Tables 2–4, component coverage and limitations on page 8; and architecture, inputs and optimization details in Appendix A on page 12. Cross-checked Figure 2's Distill arrow against Eq. (9) and Figure 3's filled/empty cells against its caption and routing paragraph. No unresolved claim-relevant conflict was identified; register-to-future routing is asymmetric and the teacher target alone is detached. All six final original crops were individually viewed, including the corrected architecture crop. All four text chunks, including reference-only pages 9–11, were read. No separate supplement, code or external source was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 12. 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
- Abstract (p. 1)
- 1 Introduction (pp. 1–3)
- 2 Related Work (p. 3)
- 3 Method, including 3.1–3.3 and Eqs. (1)–(10) (pp. 4–7)
- 4 Experiments, including 4.1–4.4 (pp. 7–8)
- 5 Limitations and Discussion; 6 Conclusion (p. 8)
- References (pp. 9–12)
- A Implementation Details (p. 12)
Outside the original text pass
- Identity/version note: the inspected title and all six authors match the catalog. The title-page margin identifies arXiv:2608.11605v1, dated 12 August 2026, whereas the catalog submittedDate is 11 August 2026. No other revision was supplied or compared; the report describes this fingerprinted v1 artifact (e1).
- 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's image limitation was addressed by inspecting original PDF pages 1–8 and 12, including all three figures and four tables. Reference-only pages 9–11 were read as text.
- No separate supplement was supplied. Code was not inspected and experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e1PDF p. 1, title, author/affiliation block and arXiv version margin
The title and six authors match the catalog after name-order normalization. The margin identifies arXiv:2608.11605v1 [cs.AI], 12 Aug 2026. Affiliations are Shanghai Jiao Tong University, ACE Robotics and Nanyang Technological University.
Go to primary source ↓e2PDF pp. 1–3, Abstract and Section 1; p. 2, Figure 1 and caption
The authors contrast cascaded, joint and direct-policy WAMs, and propose exposing hidden predictive context without future-video generation. ForeWAM-Flash uses OneDP to reduce action-denoising steps; no embodied robot-data pretraining is used.
Go to primary source ↓e3PDF pp. 4–5, Figure 2 and caption, Section 3.1, Eqs. (1)–(5)
The policy conditions on observation, language and proprioception. At deployment a clean current latent is concatenated with Gaussian future slots; a single video prefill returns dynamics-register states and layer-wise K/V context for direct action prediction.
Go to primary source ↓e4PDF p. 5, Section 3.2, Token groups and routing; p. 6, Figure 3 and following paragraph
The mask depicts current-frame, dynamics-register, future and action groups. Reading rows as queries, current tokens read current tokens; registers read current/register tokens; future rows add temporal access; action rows read all groups. Action queries concatenate cached video K/V with current action K/V.
Go to primary source ↓e5PDF p. 6, Section 3.2, Future-KV prefill and inference paragraph
At inference the video branch runs once at sigma = 1.0, retaining the clean current latent and filling future positions with pure noise. Per-layer K/V tensors are reused during action denoising; action-loss gradients remain connected to the prefill during training.
Go to primary source ↓e6PDF p. 6, Latent-action-supervised dynamics registers; p. 7, Eq. (9) and following paragraph
A frozen inverse-dynamics latent-action teacher produces a quantized, non-executable transition target. A trainable projection of mean-pooled dynamics registers matches it; stop-gradient applies only to the teacher target. The teacher is absent at deployment.
Go to primary source ↓e7PDF pp. 6–7, Section 3.3, Eqs. (6)–(10)
Video and action branches use continuous flow matching with interpolation (1-t)y+t epsilon and target velocity epsilon-y. The total loss sums video flow, action flow and weighted latent-action register matching. This objective alone does not prove causal necessity of the registers.
Go to primary source ↓e8PDF p. 7, Section 4.1, Benchmarks and evaluation protocol
Standard LIBERO uses Spatial, Object, Goal and Long suites with 50 rollouts per task. LIBERO-Plus perturbs viewpoint, robot initial state, language, lighting, background texture, sensor noise and object layout.
Go to primary source ↓e9PDF p. 7, Table 1, ForeWAM, ForeWAM-Flash and Fast-WAM rows; Section 4.2
Standard LIBERO overall success is 96.7% for ForeWAM and 96.9% for Flash versus 97.6% for Fast-WAM. ForeWAM's Spatial/Object/Goal/Long rates are 97.0/99.6/97.2/92.8%; Flash's are 97.8/99.2/97.4/93.0%. Both variants have 2B parameters versus Fast-WAM's 6B; all three show no embodied pretraining.
Go to primary source ↓e10PDF p. 8, Table 2, Ours, Ours-Flash, Fast-WAM and pi0.5 rows; p. 7, Robustness on LIBERO-Plus
On the observed subset, overall success is 61.6/58.2/51.5% for Ours/Flash/Fast-WAM. Camera success is 62.5/57.9/16.4%; noise is 58.8/53.7/37.7%; robot is 37.4/40.4/44.5%; light is 74.1/71.0/78.2%. The pi0.5 row reports 85.7% overall. External results have different sources, limiting causal comparison.
Go to primary source ↓e11PDF pp. 7–8, Section 4.3; p. 8, Table 3, all rows
Standalone action-generation latency on one NVIDIA A800 80 GB GPU is 667 ms for Fast-WAM, 568 ms for Ours and 220 ms for Flash. ForeWAM uses ten action steps; Flash uses two. These measurements are not task-completion times or rollout statistics.
Go to primary source ↓e12PDF p. 8, Section 4.4 and Table 4, all configurations
Both components yield 61.6%, Future-KV only 58.5% and LA supervision only 58.0%, each over 1,482 coverage-matched observed evaluations. The base policy reports 53.6% over 10,027 evaluations with different coverage and is explicitly contextual.
Go to primary source ↓e13PDF p. 8, Sections 5–6
Evaluation is restricted to LIBERO and LIBERO-Plus. Different morphologies, broader tasks and real-world deployment remain unverified. The authors restrict robustness conclusions to the evaluated subset.
Go to primary source ↓e14PDF p. 12, Appendix A, Architecture and inputs
The visual branch retains Wan2.1-T2V-1.3B's Video DiT, text encoder and VAE; instruction embeddings are precomputed. Both transformer branches have 30 blocks, with widths 1536 and 1024. The action expert starts from a linearly interpolated Wan2.1 ActionDiT checkpoint. H=32; 33 observation frames map to nine video frames at temporal ratio four; two 224-by-224 views are concatenated to 224-by-448. Proprioception is 8D and actions are 7D.
Go to primary source ↓e15PDF p. 12, Appendix A, Optimization and inference; p. 7, Eq. (10)
The configuration uses 16 dynamics registers, a 32D frozen LaWAM teacher target, no readability registers, a 1,000-timestep flow schedule shifted by 5.0 and AdamW with learning rate 1e-4, weight decay 0.01, cosine annealing and clipping at 1.0. No batch size, training duration/hardware count, software versions or numerical latent-action loss weight is supplied here.
Go to primary source ↓8.5 Primary sources
Foresight Without Seeing: Latent Futures for World Action Models ↗
PDF · 6,376 extracted words
Source fingerprint
8324ff119b20e43d0c2282aa6e52d17ab989546a8386badadaa39d4d3bb6e083