Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination
1. Paper overview
In one sentence: A compact video expert can guide robot actions with coarse futures and sparse updates, but the fastest configuration sacrifices some task success and requires careful latency accounting. e02e03e04e05e07e17
| At a glance | What to know |
|---|---|
| Research problem | |
| Core mechanism | |
| A key reported result | Astribot S1 four-task physical manipulation: RT: 66.25% and 98 ms on RTX 4090. Average success (%) and policy latency (ms/chunk). Dedicated policies; the same 100 demonstrations per task for all three methods; 20 trials per task; binary success within three minutes. Motus: 63.75%, 3215 ms; pi0.5: 53.75%, 113 ms. RT succeeds on 95% tray grasping, 75% bottle transfer, 65% LEGO sorting and 30% pen uncapping. The paper describes a 32× latency speedup over Motus; the small success difference lacks reported uncertainty. e07e15e17 |
| Reading caution |
Core contributions
Figure 2. Compression changes the video expert, future-token budget and frequency of video updates. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left side as training-data preparation: the first frame stays detailed while future frames pass through downsampling before VAE encoding. The center's downward arrow transfers a 30-block teacher into a 12-block student; the scissors denote structural reduction. On the right, follow inference upward from noisy actions toward the smooth action trace. The action column remains active while gray video regions indicate KV-cache reuse. The labels distinguish a 0.8B video expert from a 0.2B action expert. Section 3.2 supplies the interaction omitted by this schematic: action tokens attend to video tokens at each MoT layer, with language supplied through cross-attention. e03e04e12e14e20
What it supports. The design preserves a future-conditioning pathway while reducing its cost. The current observation retains detailed spatial information; predicted futures occupy a smaller representation. The paper therefore tests whether a compact dynamic context can guide actions, rather than requiring the action expert to depend on a fully rendered future video.
Where the evidence stops. The drawing separates refresh blocks with cached intervals; Section 3.4 gives the initial two updates as its example and Appendix E emphasizes early updates. Treat refresh placement as schematic, not an exact algorithm. Appendix A confirms that video freezing applies to Stage 2, followed by joint refinement.
2. Motivation
2.1 The problem and the proposed response
Photorealistic future generation makes world-action policies expensive. The paper asks whether a smaller future representation can retain useful control context while reducing backbone size, token count and video sampling depth. e02e03
2.2 What this reading follows
Efficient-WAM asks how much future imagination a robot policy needs to remain useful. Its answer is an asymmetric allocation of resources: preserve detail in the current observation, compress the predicted future, and spend more sampling steps refining actions than refining video. The paper separates a compact structural baseline from an RT deployment variant, a distinction essential to interpreting its results. This reading follows the transfer of video knowledge into a small expert, the evidence for reducing visual computation, and the physical controller that executes the predictions. It also preserves the source's protocol differences, unspecified implementation details and diagram–text ambiguities. e02e03e04e05e07e17
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
Insufficient evidence to decide
Separate video/action experts support a Dual-system interpretation. Equation (1) makes the action branch future-conditioned and thus inverse-dynamics-like, but the implementation jointly processes video and actions through MoT with flow-matching losses. Q4/IDM is a plausible functional reading, not an unambiguous standalone-IDM architecture. Joint training alone also does not establish One Model. e02e03e13e14
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 Spend visual capacity where the policy needs it
Start with the conditional distribution in Equation (1). The policy receives an observation, instruction and robot state; future latents provide additional context for generating an action chunk. Efficient-WAM allocates detail unevenly because present-state localization and imagined dynamics play different roles. Training future frames are downsampled before VAE encoding, while the current observation keeps higher resolution. At inference, the future branch predicts latents from noise, rather than receiving the real future. Action tokens then read that context through MoT attention. A reader interpretation is that the architecture preserves a detailed anchor for current geometry while economizing on anticipated change. Table 3(b) supports the computational benefit but also records a success penalty, so this allocation is an empirical compromise rather than a guarantee that fine visual information is irrelevant. e02e03e04e09
5.2 Separate knowledge transfer from the sampling schedule
The small video expert is not simply trained from scratch. Stage 1 starts from selected WAN layers and reduced width, then combines video flow matching with hidden-state and temporal-motion distillation. The teacher remains frozen and its supervision weight eventually decays to zero. Stage 2 attaches the action expert while freezing video parameters; Stage 3 jointly refines both experts. Independently sampled video and action noise times let the coupled forward pass encounter different corruption levels in the two streams. This training choice is distinct from the training-free inference schedule: at deployment the action branch continues refinement while cached video features replace repeated video computation. Figure 4 tests the sampling budget, whereas Table 3(a) tests transfer into the compact backbone. Neither experiment separately identifies which distillation loss preserves contact information. e04e08e10e12e13e14
5.3 Translate a policy-call speedup into the robot loop
A 98 ms policy call is only one component of physical control. Appendix C times the prediction of a 16-step chunk after warm-up with cached text embeddings, excluding sensing and robot motion. Appendix B then describes executing four or five uniformly sampled actions, each taking 0.3 seconds, before the next plan. Thus the reported amortized milliseconds per predicted step do not describe an actuator update rate or a complete feedback cycle. The hardware comparison needs equal care: Table 5 evaluates resolution-only and asymmetric-only changes independently on A800, then reports all optimizations together on RTX 4090. A reader should use the A800 rows to understand isolated changes and Table 2 to assess the deployed profile. The supplied evidence does not cleanly factor hardware, implementation and algorithmic gains into separate multiplicative speedups. e05e07e16e17e21
5.4 Training and inference
During training
Stage 1 trains the video student with video flow matching and frozen-teacher supervision. Hidden-state cosine alignment uses 256-dimensional projections; motion alignment compares spatially averaged inter-frame deltas. Distillation weight decays 0.2 → 0.1 → 0. e12e13
Stage 2 freezes the video backbone and trains actions; Stage 3 updates both experts. Action and video noise timesteps are sampled independently. Joint loss weights actions at 1.0 and video at 0.01. A video loss listed during Stage 2 must not be read as updating frozen video parameters. e04e12e13e14
All stages use AdamW, cosine scheduling, bf16, weight decay 10^-3 and reported batch size 16 × 8. Learning rates are 5 × 10^-5 initially, then 10^-5 for video and 5 × 10^-5 for actions in Stage 3. Each stage lasts approximately 2.5 simulation epochs or 5 physical-task epochs. e12e14
During inference
The structural baseline retains high-resolution futures and symmetric sampling. RT adds low-resolution futures and a fixed [T_v,T_a] = [2,10] video/action budget, caching video features during further action refinement. Figure 2 depicts separated refresh blocks, whereas the text describes early updates; exact refresh indexing is unresolved. e03e04e05e20
On Astribot S1, the controller executes four or five steps uniformly sampled from each 16-step prediction, each lasting 0.3 seconds, then replans. Policy-call latency is consequently distinct from feedback-cycle duration. e16
5.5 Implementation flow
- Transfer the video expert
Slice teacher layers [1,2,4,6,8,11,14,17,20,23,26,30] into a 12-layer student with 2048 hidden dimensions, 8192 FFN dimensions and 16 attention heads. Distillation aligns projected hidden states and frame-to-frame changes. e12e13
- Encode a multiscale context
Keep current observations detailed; downsample target future frames before VAE encoding. Illustrative sizes are 384 × 320 and 192 × 160. Patchify and concatenate both latent groups. Language enters through cross-attention; state and noisy actions become action tokens. e03
- Condition trajectory refinement on imagination
Action tokens attend to video tokens at MoT layers. Future prediction therefore participates in inference-time control, beyond auxiliary training supervision. Predicted chunks still require a separate robot execution loop. e02e03e04e16
6. Experiments & results
Efficient-WAM compresses a future-video expert and uses its coarse latent predictions to guide action generation. Its RT variant combines sparse future tokens with fewer video updates, trading some simulation success for faster policy calls. The evidence supports an efficiency–control tradeoff, without establishing that visual detail is universally dispensable.
6.1 Read the original evidence
Table 1. The compact baseline and RT variant occupy different points on the success–efficiency tradeoff. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the final two rows: both share the 1B parameter entry, but only RT adds the future-token and sampling reductions. Read Clean and Random as separate evaluation settings, not as uncertainty bounds. The authors' policies are evaluated on 50 tasks with 100 trials per task per setting, after training on clean and randomized demonstrations. Moving upward provides the reported baseline comparisons; the method groups organize model families rather than establish identical training histories. Appendix E is essential context: the authors evaluate their own checkpoints, while baseline task scores are primarily drawn from existing publications. e05e06e19
What it supports. Efficient-WAM reports 86.7% clean and 85.7% randomized success, compared with RT's 83.1% and 82.0%. The RT changes therefore cost 3.6 and 3.7 percentage points. The compact baseline approaches Motus's 88.7% and 87.0% with a smaller reported parameter count, but it does not surpass Motus here.
Where the evidence stops. This table does not supply latency or confidence intervals. Cross-paper baseline provenance limits a fully controlled ranking. Keep its 100-trial evaluation distinct from the 20-rollout ablations; their absolute success rates cannot be treated as paired measurements.
Table 2. The deployed RT model reports 66.25% average success and 98 ms per policy call. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Here Ours means Efficient-WAM-RT. Read the first four rows as binary task-success percentages, followed by average success and two timing summaries. Every method uses the same 100 demonstrations per task, a dedicated policy for each task and 20 evaluation trials per task. Success has concrete physical requirements: keeping the tray level and untouched on top, avoiding bottle collisions, sorting every LEGO block, or removing a pen cap without drops or tipping the holder. The three-minute trial cap and randomized feasible object resets belong to this protocol. Read the timing rows with Appendix C's policy-call definition, not as motion durations. e07e15e16e17
What it supports. RT averages 66.25%, compared with Motus at 63.75% and pi0.5 at 53.75%. Its 98 ms chunk latency is far below Motus's 3215 ms; the authors summarize this as a 32× speedup. Performance remains task-dependent: RT reaches 95% on tray grasping but only 30% on pen uncapping.
Where the evidence stops. The small mean-success advantage has no reported uncertainty. The 6.1 ms per-step entry amortizes chunk computation; it is not physical feedback timing. Deployment executes four or five sampled actions, each lasting 0.3 seconds, before replanning.
Figures 7–8. The RT futures visibly lose detail, providing qualitative context for the control–efficiency tradeoff. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The upper panel is Figure 7's Full WAN configuration; the lower panel is Figure 8's Efficient-WAM-RT. Within each panel, use the Condition frame and Prediction frame labels and follow the horizontal arrow to the right. Compare matching Adjust Bottle and Grab Roller strips across the two configurations. Full WAN produces more distinct gripper and object boundaries, while RT shows blur, ghosting and changed textures. Do not treat the image sequence as a physical rollout: these are predicted future frames. Appendix D ties the visual difference to the compact expert, low-resolution future latents and asymmetric updates operating together. e18e05e06e08
What it supports. The examples make the intended reduction in visual fidelity directly inspectable. Recognizable scene and gripper structure can coexist with severe artifacts. Together with the task results, this supports the narrower finding that strong control is possible with visibly coarse futures in the evaluated settings; it does not show which visual features the policy actually uses.
Where the evidence stops. These selected examples have no quantitative fidelity metric, ground-truth error analysis or causal feature test. Appendix D's Full-WAN-versus-RT numerical comparison draws from different rollout protocols in Tables 3 and 1, so the pictures cannot establish a controlled fidelity–success curve.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0 clean and randomized manipulation 50 tasks; 2,500 clean and 25,000 randomized demonstrations; one multitask policy; 100 trials per task per setting. | Efficient-WAM: 86.7 clean / 85.7 randomized; RT: 83.1 / 82.0. Average task success (%) | Motus: 88.7 / 87.0; LingBot-VLA: 86.5 / 85.3; GigaWorld-Policy: 86.4 / 85.0. RT loses 3.6/3.7 percentage points versus the compact baseline. Appendix E identifies baseline scores as largely imported from publications, limiting claims of fully matched retraining. e05e06e14e19 |
| Astribot S1 four-task physical manipulation Dedicated policies; the same 100 demonstrations per task for all three methods; 20 trials per task; binary success within three minutes. | RT: 66.25% and 98 ms on RTX 4090. Average success (%) and policy latency (ms/chunk) | Motus: 63.75%, 3215 ms; pi0.5: 53.75%, 113 ms. RT succeeds on 95% tray grasping, 75% bottle transfer, 65% LEGO sorting and 30% pen uncapping. The paper describes a 32× latency speedup over Motus; the small success difference lacks reported uncertainty. e07e15e17 |
| Compact expert and future-resolution ablations RoboTwin, 20 rollouts per task across 50 tasks; A800 latency measurements. | Random student: 69/68; sliced: 82/81; sliced plus distillation: 87/86. High-to-low future resolution: 87/86 → 83/82, 240 → 60 tokens, 430 → 377 ms. Clean/randomized success (%) and ms/chunk | Full WAN: 86.4/85.5 at 2013 ms; distilled compact model: 430 ms. Initialization and distillation matter at fixed student size. Resolution reduction carries a four-point success loss in both settings; it does not preserve accuracy exactly. e08e09e17 |
| Asymmetric denoising on RoboTwin clean 50 tasks, 20 rollouts per task; compact baseline with high-resolution futures; A800. | [2,10]: 86.3% at 139 ms; repeat: 86.2% at 140 ms. Average success (%) and ms/chunk | [10,10]: 87.1%, 431 ms in Figure 4 (430 ms in text); [1,10]: 79.3%, 92 ms. Two video updates preserve much more performance than one. These are independent baseline ablations, not successive stages on the way to the RTX 4090 result. e10e17e20 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3(a–b). Knowledge transfer recovers much of the compact expert's performance; fewer future tokens incur a measurable success loss. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read panel (a) down the compact rows to separate initialization from distillation. Student video size stays at 0.8B, while Random/No, Sliced/No and Sliced/Yes change how teacher knowledge enters training. Params here refers to the video expert, unlike the complete 1B system entry in Table 1. In panel (b), move from High through Medium to Low and track both token count and success; Lat. is milliseconds per chunk. The evaluation uses 20 rollouts on each of 50 tasks. Appendix C clarifies that the resolution and asymmetric-sampling variants are independently applied to the compact baseline, despite the main text's progressive-ablation wording. e03e08e09e13e17
What it supports. Random initialization yields 69%/68% clean/randomized success; slicing raises it to 82%/81%, and distillation to 87%/86%. Lowering future resolution reduces tokens from 240 to 60 and latency from 430 to 377 ms, while success falls to 83%/82%. The evidence supports a useful tradeoff rather than zero accuracy cost.
Where the evidence stops. Distillation combines hidden-state and temporal-motion losses, so this ablation does not isolate their individual contributions. High/Medium/Low are the table's labels; the source does not map every row to exact pixel dimensions. No confidence intervals are reported.
Figure 4. Keeping ten action updates while reducing video updates to two gives the strongest illustrated sampling compromise. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each horizontal label as [T_v,T_a], the number of video and action denoising steps. Bar height uses the left latency axis; the line uses the right success axis. Color distinguishes video budgets and bar patterns distinguish action budgets. The shaded region selects two [2,10] evaluations, with the starred point corresponding to the second run described in Table 7. To isolate video sampling, compare [10,10], [5,10], [2,10] and [1,10], keeping the action budget fixed. Other plotted points also reduce action refinement, so the full line is not a sweep of video steps alone. Appendix C places this ablation on the high-resolution compact baseline. e10e17e20
What it supports. Reducing ten video updates to two changes clean success from 87.1% to 86.3%, while latency drops from the plotted 431 ms to 139 ms. One video update is faster at 92 ms but falls to 79.3%. The steep loss below two updates motivates retaining some inference-time future computation.
Where the evidence stops. The graphic reports 431 ms for [10,10], whereas the adjacent text and Table 5 report 430 ms; both are preserved. Two [2,10] runs are not a confidence interval. These A800 timings do not measure the full RT deployment on RTX 4090.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors warn that coarse futures may hurt micro-manipulation and that fixed sampling schedules cannot allocate extra compute to uncertain dynamics. Observed failures include misalignment, missed objects and collisions. e11e16
Figures 7–8 demonstrate blur and ghosting, not measured contact accuracy or causal use of future cues. Appendix D compares RT main results with a Full-WAN ablation using a different rollout count. e08e06e18
Timing excludes instruction encoding, observation acquisition and robot motion. A800-to-RTX-4090 comparisons change hardware; Section 4.5's overhead explanation cannot establish a controlled algorithm-only gain. That section also calls the video expert 1B, whereas Figure 2/Table 3 specify 0.8B. e17e21e03e08
7.2 Questions for discussion
- Does matching policy latency change which future resolution is best?
- Would shuffling future context reduce success, demonstrating that the action branch actually uses its predicted dynamics?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs WAN-2.2-5B weights, the sliced architecture, VAE/text conditioning, RoboTwin demonstrations and the physical task data/end-effector. Preserve task-specific binary criteria and the receding-horizon loop. e03e06e12e15e16
The source leaves exact retained width indices, action-expert configuration, distillation transition points and video-refresh indices incomplete. Batch size 16 × 8 does not establish a training GPU model/count; training wall time and software versions are unspecified. e03e12e13e14
A proposed check holds hardware, checkpoint and rollout seeds fixed while varying future resolution and video updates, measuring success and policy latency separately from execution time. e09e10e17
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 the action expert use the predicted future?
Reader-proposed, not performed: keep one RT checkpoint, current-observation tokens, instruction, state, action noise and [2,10] schedule fixed. Compare normal future-token context with context shuffled across episodes of the same task and with the future contribution masked, leaving current-observation context intact. Use matched RoboTwin reset seeds and report per-task success across repeated evaluation seeds. Verify that interventions affect only future tokens. A reproducible success drop under both interventions would support causal use of future context; unchanged success would weaken the claim that the retained imagination branch supplies essential control guidance. e02e03e04e06e20
Check 2: Measure the combined speed–success tradeoff on one GPU
Reader-proposed, not performed: train matched compact students at high and low future resolution using the same data, transfer recipe and training seeds. Evaluate each with [10,10] and [2,10] on one A800, holding action horizon, execution interface and compilation settings fixed. Time policy calls after the paper's warm-up with cached text, and measure sensing/motion separately. Use the same rollout budget and reset seeds for all four cells. Report uncertainty across training/evaluation seeds. If the combined configuration offers no repeatable same-hardware latency reduction, or its success loss exceeds the separate effects, the claimed complementary acceleration would require qualification. e08e09e10e12e14e16e17
8.3 Reading coverage
Visual audit: Visually inspected the title/version/author page, Figures 1–8, Tables 1–7, method equations, training details, physical success/execution criteria, failure cases, latency protocol and appendix provenance. All supporting method, numerical and reproduction pages are declared. Reference-only pages 9–11 were read in the complete text but not rendered. Six final original crops were inspected; the denoising crop was corrected and inspected again to retain its left axis. The combined future-prediction crop preserves the original stacked layout and short intervening Figure 7 caption. Figure 2's schematic refresh placement and Figure 4's 431-versus-430 ms discrepancy remain explicit. Separate supplemental material was not supplied or verified.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15, 16, 17, 18. 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, abstract and version stamp (p. 1)
- 1 Introduction (p. 2)
- 2 Related Works and 3.1 Design Formulation (p. 3)
- 3.2–3.5 Architecture, multiscale latents, asymmetric denoising and objectives (pp. 4–5)
- 4.1–4.5 Experiments, ablations and efficiency (pp. 5–8)
- 5 Conclusion and 6 Limitations (p. 8)
- Acknowledgments and References (pp. 9–11)
- Appendix A Training Details (pp. 12–13)
- Appendix B Real-World Evaluation Details (pp. 13–14)
- Appendix C Latency Measurement Protocol and Summary (p. 15)
- Appendix D Qualitative Future Prediction Results (pp. 15–16)
- Appendix E RoboTwin Detailed Results (pp. 17–18)
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's figure limitation was addressed by inspecting original PDF figures and tables; all five supplied text chunks were read completely.
- Identity matches the catalog title and all twelve authors. The inspected artifact is arXiv:2606.10040v2 [cs.RO], stamped 10 Jun 2026. The catalog URL is unversioned; no earlier version was supplied, so revision differences cannot be assessed.
- No separate supplement was supplied. Code and external project resources were not inspected, and experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block and arXiv margin stamp
Exact catalog title and all twelve authors match; the inspected source is arXiv:2606.10040v2 [cs.RO], 10 Jun 2026. Five affiliations are printed.
Go to primary source ↓e02PDF p. 3, Section 3.1, Equations (1)–(2)
The joint distribution factorizes into future-video and future-conditioned action branches; video cost depends on model size, token resolution and denoising budget.
Go to primary source ↓e03PDF p. 4, Figure 2/caption and Sections 3.2–3.3
Figure labels video/action experts 0.8B/0.2B and teacher/student depths 30/12. Text specifies cross-attention, state/action tokens and multiscale VAE latents. The inference diagram separates refresh blocks with cached intervals.
Go to primary source ↓e04PDF p. 5, Sections 3.4–3.5, Equation (3)
Video features are cached between fewer video updates, with the initial two given as an example. Flow matching interpolates Gaussian noise to clean targets; training uses video adaptation, frozen-video action training and joint refinement.
Go to primary source ↓e05PDF p. 5, Section 4.1
Baseline uses high-resolution futures and symmetric denoising; RT combines sparse future latents and asymmetric sampling. H = 16; simulation uses A800 and physical deployment RTX 4090.
Go to primary source ↓e06PDF p. 6, Section 4.2 and Table 1, method/Clean/Random/Params columns
50 tasks, 2,500 clean plus 25,000 randomized demonstrations, 100 trials per task/setting. Baseline 86.7/85.7 and RT 83.1/82.0; Motus 88.7/87.0, LingBot-VLA 86.5/85.3, GigaWorld-Policy 86.4/85.0.
Go to primary source ↓e07PDF p. 6, Section 4.3 and Table 2, all rows
Equal per-task physical training/evaluation counts; RT success 95/75/65/30 and mean 66.25, versus Motus 63.75 and pi0.5 53.75. Latencies are 98/3215/113 ms respectively.
Go to primary source ↓e08PDF p. 7, Section 4.4 and Table 3(a), all variant rows
Ablations use 20 rollouts/task. Random, sliced and distilled students score 69/68, 82/81 and 87/86. Full WAN scores 86.4/85.5 at 2013 ms; distilled student is 0.8B at 430 ms.
Go to primary source ↓e09PDF p. 7, Table 3(b), High/Medium/Low rows
High/medium/low futures use 240/126/60 tokens, score 87/86, 85/84, 83/82, and take 430/396/377 ms.
Go to primary source ↓e10PDF p. 8, Figure 4/caption and Section 4.4 continuation
Figure plots [2,10] at 86.3%,139 ms and repeat at 86.2%,140 ms; [1,10] at 79.3%,92 ms. [10,10] is 87.1%,431 ms in the graphic but 430 ms in prose.
Go to primary source ↓e11PDF p. 8, Section 6 Limitations
The authors identify possible precision costs of low-resolution futures and static asymmetric schedules, proposing dynamic allocation as future work.
Go to primary source ↓e12PDF p. 12, Appendix A, Table 4 and architecture paragraph
Specifies stage trainability, batch 16 × 8, learning rates, loss weights, optimizer, precision, weight decay, 12 selected teacher layers, hidden/FFN widths and attention heads.
Go to primary source ↓e13PDF p. 12, Appendix A, Equations (4)–(7) and surrounding text
Frozen-teacher supervision uses projected hidden-state and temporal-delta cosine alignment. Distillation weight decays 0.2→0.1→0. Video/action noise times are independently sampled.
Go to primary source ↓e14PDF p. 13, Appendix A continuation, Equation (8) and training-duration paragraph
Joint loss combines action and video velocities with stage-specific weights; simulation trains one multitask policy and physical tasks separate policies, for about 2.5/5 epochs per stage.
Go to primary source ↓e15PDF p. 13, Appendix B, evaluation protocol and task criteria
Astribot S1 receives three RGB views and 31-dimensional joint states; 100 demonstrations and 20 trials/task, randomized feasible resets and three-minute cap. Tray, bottle and LEGO success criteria are specified.
Go to primary source ↓e16PDF p. 14, Appendix B, pen criterion, execution protocol and Figure 6/failure analysis
Controller executes four or five uniformly sampled steps from H=16, each 0.3 s, before replanning. Failures include misalignment, missed objects and collisions; pen success also requires no drops or tipped holder.
Go to primary source ↓e17PDF p. 15, Appendix C and Table 5, all configuration rows
Policy-call timing follows one warm-up with cached text, excluding instruction encoding, sensing and motion. Compact, low-resolution-only and asymmetric-only A800 profiles are 430/377/139 ms; full RT is 98 ms on RTX 4090.
Go to primary source ↓e18PDF p. 15, Appendix D; PDF p. 16, Figures 7–8 and captions
Full WAN and RT future examples cover Adjust Bottle and Grab Roller. RT predictions show blur/ghosting; the text links them to results in Tables 1 and 3 without a separate visual-fidelity metric.
Go to primary source ↓e19PDF p. 17, Appendix E, Table 6 and provenance paragraph
Own checkpoints use 100 rollouts/task; baseline per-task values primarily come from other publications, including pi0/pi0.5 results sourced through LingBot-VLA.
Go to primary source ↓e20PDF p. 18, Table 7, Average row and explanatory paragraph
Clean denoising sweep uses 20 rollouts/task and two [2,10] runs (86.3/86.2). The text locates video updates in the early high-noise regime and reports losses from single-video-step/action compression.
Go to primary source ↓e21PDF p. 8, Section 4.5 System-Level Efficiency Analysis
The narrative calls the video expert 1B and attributes deployment gains partly to removal of simulation overheads. Figure 2/Table 3 instead label the video expert 0.8B; Appendix C supplies the more precise timing boundary.
Go to primary source ↓8.5 Primary sources
Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination ↗
PDF · 8,898 extracted words
Source fingerprint
03bceaba02b236070153c14f956db378efa4aa2c8d3aaf6df6d3a70be16073d7