Flash-WAM: Modality-Aware Distillation for World Action Models
1. Paper overview
In one sentence: Matching consistency functions to video and action noise regimes makes a shared world-action model much faster, with lower task success and unresolved baseline-reporting conflicts. E02E04E05E06E07E13E14E15
| At a glance | What to know |
|---|---|
| Research problem | Source description LingBot-VA first predicts future video latents, then denoises actions conditioned on that future. Both stages cost sequential transformer evaluations. Uniform consistency distillation ignores their different noise distributions, potentially weakening action learning even when video remains plausible. E02E03E04 |
| Core mechanism | |
| A key reported result | RoboTwin 2.0 manipulation at 1v/1a: 82.56 Clean; 80.26 Randomized; 81.41 average; reported 23.3× speedup. Task success (%) and per-chunk latency. Same 50-task benchmark; single NVIDIA L40S latency profiling. Teacher: 91.25% at 25v/50a and 8,100 ms. Student: 348 ms in prose, 347 ms in Figure 1. The fastest setting loses 9.84 percentage points. Both reported times meet the authors’ 500 ms budget; the one-millisecond discrepancy remains unresolved. E07E06 |
| Reading caution |
Core contributions
Figure 2. Different consistency maps train one shared video-action backbone. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left panel vertically: the upper curves show where video and action encounter noise, while the lower curves show how the velocity coefficient changes with that noise. Flash-WAM’s action line has magnitude σ, whereas the LCM curve approaches zero more rapidly. In the middle, a teacher Euler step moves from σ_s to lower σ_e; student and EMA-target endpoint estimates are compared using separate video and action losses. Their sum includes the action weight λ_a. At deployment, follow the right panel’s shared backbone and autoregressive loop, interpreting video-then-action order through Eqs. (5)–(6). Its bottom label permits one or two action steps. E02E03E04E05
What it supports. The mechanism changes how each stream is supervised, while retaining the shared transformer and explicit future-video generation. The action map directly addresses the low-noise coefficient attenuation diagnosed by the authors. The graphic supports this architectural account; it does not itself demonstrate improved executed-action accuracy or a guaranteed gradient norm.
Where the evidence stops. The orange action arrows enter f^v, and f^v points to f^a; this routing is ambiguous against the separate maps in Eqs. (9)–(11). The explanation follows those equations. Also, Na=1/2 in the graphic is broader than the caption’s single-step wording. Neither discrepancy is silently corrected.
2. Motivation
2.1 The problem and the proposed response
LingBot-VA first predicts future video latents, then denoises actions conditioned on that future. Both stages cost sequential transformer evaluations. Uniform consistency distillation ignores their different noise distributions, potentially weakening action learning even when video remains plausible. E02E03E04
2.2 What this reading follows
Flash-WAM asks why a successful video-distillation recipe can fail when the same network also predicts robot actions. Its answer concerns where each modality encounters noise during training: video emphasizes high noise, while actions also need useful learning signals near clean data. The proposed fix keeps LingBot-VA’s shared transformer and its video-conditioned action generation, but changes the action consistency map. Read the figures as a linked argument about training signal, sampling cost and measured task success. The strongest headline success rates use two action steps; the fastest setting uses one. Several appendix and table discrepancies limit how confidently the baseline comparisons can be interpreted. E02E04E05E06E07E13E14E15
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
WAM, efficient inference and IDM are supported: actions depend on predicted visual transitions. Both stages explicitly share transformer parameters, and Flash-WAM changes loss parametrizations. This supports One Model × IDM, conflicting with the recorded Dual-system architecture and Q4 quadrant. The catalog snapshot remains unchanged. E02E05E15
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 Start with the conditional policy, not the loss diagram
The source defines two conditional distributions: future video given context, followed by actions given that future and the same context. Context includes earlier observations, actions and the language instruction. This is inverse dynamics because the action generator conditions on a proposed visual transition. It is also a shared model: both stages use the same transformer parameters. Joint training concatenates video and action tokens, but inference still orders their generation. Reducing denoising steps therefore shortens both parts of an existing pipeline; it does not remove future prediction or replace actions with a video. This distinction explains why the catalog’s IDM label fits while its Dual-system architecture label conflicts with the reviewed implementation. The training teacher and EMA target are not extra deployed controllers. E02E03E05
5.2 Separate noise coverage from coefficient attenuation
A consistency function maps a noisy point on a teacher trajectory toward a clean endpoint. In the paper’s general form, the network velocity is multiplied by a noise-dependent coefficient b. Standard LCM makes its magnitude vanish quadratically near zero; Flash-WAM’s action map uses b=−σ, which vanishes linearly. Because the action schedule places substantial mass at low noise, the authors argue that this gives actions a more useful learning signal. Video retains LCM because its schedule emphasizes higher noise. Reader analysis: this isolates a coefficient in the chain rule, not the complete gradient, whose size also depends on prediction residuals and network derivatives. The proof’s stated smoothness assumption and printed expansion need correction before accepting its full mathematical wording. E03E04E05E12
5.3 Read every success rate together with its denoising budget
The speed claim and the highest success claim refer to different operating points. On RoboTwin, one video plus two action steps gives 85.54% average success and a reported 19-fold speedup. One plus one gives 81.41% and the fastest reported latency. LIBERO likewise has separate two-action and one-action rows, and the robot average falls from 60% to 50% when moving between those settings. Reader analysis: these are operating-point tradeoffs, not evidence that compression leaves performance unchanged. Next examine whether each comparator actually used the labeled budget. Video-only LCM’s appendix specifies fifty action steps despite reduced-step table labels, and DMD2’s variant identity conflicts across tables. Those issues should be resolved before turning the reported rankings into a definitive algorithm comparison. E06E07E08E09E13E14
5.4 Training and inference
During training
LIBERO teachers are fine-tuned separately per suite for 4,000 steps with AdamW, learning rate 1e-5 and effective batch 120, then distilled for 2,000 steps. Each stage is described as approximately 24 hours per suite on four H100 GPUs. E11
Table 6 specifies 128×128 images, 30 action dimensions, four actions/frame and K=4 frames. Distillation uses video/action shifts 5/1, action weight 1, EMA decay 0.995, data scale 0.5, Huber c=0.001, video guidance [2,10], AdamW learning rate 5e-6 and effective batch 48. E12
During inference
The deployed student retains video-then-action generation at 1v/1a or 1v/2a: one video step followed by one or two action steps. Predicted futures remain inference-time conditioning. Teacher and consistency losses are training machinery; exact observation-refresh and action-execution timing is incompletely documented. E02E05E06E09
5.5 Implementation flow
- Retain shared parameters and ordered generation
Visual dynamics and inverse dynamics share transformer parameters. During joint training, concatenated video/action tokens use flex attention; modality-specific loss heads leave the backbone and per-step compute unchanged. E02E05
- Match endpoint maps to noise regimes
Independent SNR shifts concentrate video at higher noise; action has substantial low-noise mass. Action uses the clean flow estimate directly, producing a velocity coefficient linear in noise. Video retains LCM coefficients, motivated by high-noise stability. E03E04E05
- Distill teacher trajectories
A frozen teacher takes an Euler step from higher to lower noise. The student endpoint estimate matches an EMA-target estimate at the lower-noise point. Teacher classifier-free guidance applies to video only. The two consistency losses are combined. E03E05
6. Experiments & results
Flash-WAM accelerates LingBot-VA with different video and action consistency functions while retaining a shared transformer and video-conditioned inverse dynamics. It trades some task success for lower denoising cost. The headline 85.54% RoboTwin result uses one video and two action steps; the fastest one-plus-one configuration reaches 81.41%. Internal protocol and table conflicts qualify the comparisons (E02, E05, E07, E14, E15).
6.1 Read the original evidence
Figure 1. Faster sampling crosses the chosen latency budget while sacrificing success. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the blue video and orange action segments on the left. The teacher’s 25v/50a row totals 8,100 ms, so accelerating only its video stage leaves a large action cost. The highlighted Flash-WAM 1v/1a row prints 193 ms for video and 154 ms for action. The dashed red marker represents the authors’ approximately 500 ms chunk budget. Then move to the right panel and read the action-step labels beneath each bar: Flash-WAM’s 85.5 and 81.4 belong to different budgets. The teacher’s dashed horizontal line remains above both. Use Table 1 for the more precise success percentages. E06E07
What it supports. The figure reports a 347 ms single-video/single-action chunk, compared with 8.1 seconds for the teacher on an L40S. This meets the authors’ chosen budget. However, the fastest success bar is lower than the two-action-step bar and the teacher, so the result demonstrates a speed–success tradeoff rather than unchanged accuracy.
Where the evidence stops. The prose reports 348 ms rather than the graphic’s 347 ms. The 85.5% bar uses two action steps. A chunk-level profiling threshold does not establish end-to-end closed-loop timing or physical control stability.
Table 1. The two reduced-step groups must be compared separately. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use Nv and Na to identify the sampling budget before comparing any success number. The top block contains reference policies and the authors’ reproduced teacher, marked with an asterisk. The middle group uses one video and two action steps; the bottom group uses one of each. Clean and Randomized are different evaluation conditions across fifty tasks, and Average summarizes them. Read the shaded Flash-WAM row against the teacher to assess retained performance, then against methods within its own budget group to examine distillation. The speedup column is shared across each reduced-step group; the table alone does not resolve the conflicting baseline implementations described in the appendix. E07E09E10E13E14
What it supports. Flash-WAM reaches 85.54% at 1v/2a and 81.41% at 1v/1a, compared with the teacher’s 91.25%. Those are losses of 5.71 and 9.84 percentage points. The reported speedups are 19.0× and 23.3×. The table supports substantial recovery relative to naive joint LCM, while preserving a measurable teacher–student gap.
Where the evidence stops. Appendix A.3.2 gives Video-only LCM fifty action steps, contradicting these labels. The 1v/1a DMD2 row matches Appendix B’s Joint DMD2 despite a different textual assignment. Table 4 also reports different teacher and naive-LCM values; no intervals are provided.
Table 2. High average success conceals a larger Object-suite drop. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read horizontally across the four suite columns before relying on Average. Spatial, Object, Goal and Long-horizon are separate LIBERO suites, each described with five hundred demonstrations; the appendix fine-tunes a teacher separately for each suite. The asterisk marks the teacher result reproduced by the paper’s authors. Notice that this teacher uses twenty video steps and fifty action steps, unlike RoboTwin’s video budget. Compare the two shaded Flash-WAM rows to see what changes when the action budget falls from two to one. Finally, check the Object column against the teacher: it exposes a larger loss than the overall average suggests. E08E11E13
What it supports. Table 2 reports Average success of 95.7% at 1v/2a and 95.1% at 1v/1a, below the teacher’s reported 98.6%. The 1v/2a aggregate has an unresolved arithmetic conflict described below. At two action steps, Object success is 92.8% versus 99.8% for the teacher, showing that degradation is not uniform across the four suites.
Where the evidence stops. The printed 1v/2a suite values 97.0, 92.8, 96.4 and 98.0 have an unweighted mean of 96.05%, not the table’s 95.7%. The source does not explain a weighting or other reconciliation, so 95.7 is retained as reported rather than silently replaced. Video-only LCM’s action-step protocol remains inconsistent with Appendix A.3.2. The 16.3× speedup is reported as printed; the prose’s 6,767-to-404 ms values do not exactly reproduce that ratio. Uncertainty intervals are absent.
Table 3. The reported robot gain is recovery from step reduction, with a remaining teacher gap. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Decode the task abbreviations using the setup on the preceding page: T1 opens a pot and places a potato inside, T2 picks the red bottle despite a yellow distractor, and T3 moves the pink object to a marked location. The robot is a Unitree G1 with Dex1-1 grippers. Each task has fifty teleoperated demonstrations and ten evaluation rollouts per method. Read the first row as the higher-budget 3v/10a teacher, then compare methods within each reduced-budget block. The two shaded rows show Flash-WAM at 1v/2a and 1v/1a; their averages must not be merged into a single-step claim. E09E15E17
What it supports. The table reports Flash-WAM averages of 60% with two action steps and 50% with one. Direct step reduction gives 40.0% and 23.3%, while the higher-budget teacher reaches 66.7%. These are reported physical task-success measurements, distinct from the open-loop video illustrations elsewhere in the paper.
Where the evidence stops. Appendix C nevertheless states that physical deployment remains future work, contradicting this table and Figure 3. That conflict is unresolved. Ten rollouts per task and no reported confidence intervals further limit the precision of the comparison.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0 manipulation at 1v/2a 50 bimanual tasks; Clean and Randomized evaluation splits. | 88.42 Clean; 82.66 Randomized; 85.54 average; 19.0× speedup. Task success (%) and reported speedup | Table 1 teacher at 25v/50a: 91.25 average; naive joint LCM at 1v/2a: 23.97. The student loses 5.71 percentage points versus the teacher. This is the 85.5% headline setting, not single-step action generation. E07E06 |
| RoboTwin 2.0 manipulation at 1v/1a Same 50-task benchmark; single NVIDIA L40S latency profiling. | 82.56 Clean; 80.26 Randomized; 81.41 average; reported 23.3× speedup. Task success (%) and per-chunk latency | Teacher: 91.25% at 25v/50a and 8,100 ms. Student: 348 ms in prose, 347 ms in Figure 1. The fastest setting loses 9.84 percentage points. Both reported times meet the authors’ 500 ms budget; the one-millisecond discrepancy remains unresolved. E07E06 |
| LIBERO manipulation Spatial, Object, Goal and Long-horizon suites; 500 demonstrations per suite; separately fine-tuned teachers. | Table 2 reports 95.7 at 1v/2a, 13.7×; 95.1 at 1v/1a, 16.3×. Average success (%) and reported speedup | Teacher at 20v/50a: 98.6; table-labeled Video-only LCM: 95.2/94.2. The Object suite falls from 99.8 to 92.8 at 1v/2a. The four printed 1v/2a suite values, 97.0, 92.8, 96.4 and 98.0, have an unweighted mean of 96.05%, not the reported Average of 95.7%. The source does not explain a weighting or other reconciliation; 95.7 is retained as reported. Video-only LCM’s inference protocol is contradictory across the source. E08E11E13 |
| Unitree G1 manipulation Dex1-1 grippers; pot/potato, red-bottle-with-distractor and pink-object-placement tasks; 50 demonstrations and 10 test rollouts per task/method. | 60.0 at 1v/2a; 50.0 at 1v/1a. Average success (%) | Teacher at 3v/10a: 66.7; reduced-NFE teacher: 40.0/23.3; table-labeled Video-only LCM: 43.3/33.3. Reported execution success improves over direct step reduction. Small rollout counts and Appendix C’s contradictory simulation-only statement limit confidence. E09E15 |
| RoboTwin action-distillation ablation Table 4, 1v/1a, averages across 50 tasks with horizon breakdowns. | Flash-WAM 82.56/80.26; Video-only LCM + regularizer 53.48/48.40. Clean/Randomized success (%) | Table-labeled Video-only LCM: 77.90/69.46; naive joint LCM: 39.68/32.96. A demonstration-based action regularizer does not match modality-aware consistency here. Reporting conflicts prevent an unqualified all-horizons claim. E10E13E16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 4. Action-loss design matters, but the all-horizons claim exceeds the table. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First distinguish denoising steps Nv/Na from task horizon: the latter counts one, two or three sequential task steps. Within each horizon, Clean and Randomized are separate columns. Then compare the middle and lower budget blocks. Video-only LCM + reg. supervises action velocity using a masked demonstration-based flow-matching MSE, whereas Flash-WAM uses action consistency. At 1v/1a, the regularized variant has much lower average success. Before generalizing this pattern, inspect the upper block’s horizon-three Randomized column: Flash-WAM is not the largest entry. Also keep this table’s teacher row separate from Table 1 until their differing averages are reconciled. E07E10E13E16
What it supports. At 1v/1a, Flash-WAM averages 82.56/80.26% on Clean/Randomized, compared with 53.48/48.40% for the regularized variant. However, at 1v/2a its horizon-three Randomized success is 63.25%, below Video-only LCM’s 68.25%. The evidence supports a strong aggregate advantage, not superiority in every horizon and split.
Where the evidence stops. Teacher averages and naive-LCM 1v/2a Randomized values differ from Table 1. The horizon-three naive-LCM zeros also conflict with Table 8’s nonzero Stack Bowls Three row. Treat horizon conclusions as provisional until the aggregation and protocols are reconciled.
7. Analysis & limitations
7.1 What the evidence leaves open
Only one shared-backbone WAM is studied. Multi-model architectures, high-noise theory and broader distillation transfer remain future work. Success tables provide no uncertainty intervals or repeated-seed statistics. E07E08E09E15
Video-only LCM is described with 50 action steps but tabulated with one or two. Main-table DMD2 is described as video-only plus regularization, yet its 1v/1a values match Appendix B’s Joint DMD2. Appendix C defers physical experiments despite Figure 3/Table 3. E13E14E15
Table 4’s teacher averages and naive-LCM 1v/2a Randomized value differ from Table 1. Its horizon-three naive-LCM zeros conflict with Table 8. Flash-WAM loses one horizon-three Randomized comparison, contradicting the all-horizons prose. E07E10E16
The coefficient argument is narrower than a guarantee about total gradients or control. Proposition 1 invokes a quadratic Taylor remainder under only C¹ regularity, which does not justify that remainder. The printed LCM expansion’s leading coefficient also conflicts with its exact formula. E04
Figure 4 shows open-loop predicted video, not executed action precision. The 500 ms chunk budget is a profiling criterion, not a universal stability guarantee. E06E17
7.2 Questions for discussion
- Does coefficient scaling explain action accuracy after accounting for residual size and shared-backbone gradient interference?
- Which actual checkpoints and inference settings generated the inconsistent baseline rows?
- How much latency benefit survives observation processing and robot execution delays?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the correct checkpoint, suite-specific teachers, demonstration data and action normalization/masks. The PDF leaves software versions, exact teacher-step grid/skip, simulation rollout counts and full deployment timing insufficiently documented. E03E09E11E12E13
Proposed check: change only the action map between LCM and the linear form, holding teacher, data, schedules and budget fixed. Measure gradient norms by noise bin, action error and matched-seed success at both action budgets. E04E10E11E12
Proposed check: separate full-50-step and reduced-step Video-only LCM from naive joint LCM and Flash-WAM. Log actual forward passes/latency and regenerate task/horizon aggregates from identical rollouts to test the disputed comparisons. E06E07E10E13E16
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 map improve learning specifically at low noise?
Reader-proposed experiment, not performed: start both students from the same suite-specific teacher and use identical demonstrations, random seeds, noise draws, video loss and training budget. Change only the action map from LCM to the linear form. Log action-head and shared-backbone gradient norms in noise bins, held-out action error, video loss and task success at 1v/1a and 1v/2a. The hypothesis predicts a low-noise action-learning improvement without sacrificing video learning. Similar low-noise behavior and success in both conditions would weaken the claimed mechanism. E04E05E10E11E12
Check 2: Can explicit inference settings reproduce the reported tables?
Reader-proposed audit, not performed: evaluate Video-only LCM with fifty action steps and with explicitly reduced one/two-step schedules, alongside naive joint LCM and Flash-WAM. Freeze the test seeds, checkpoint selection and observation pipeline. Record actual transformer evaluations, per-chunk L40S latency and per-task success before generating horizon aggregates. Check whether Table 1 and Table 4 teacher rows can arise from the same rollouts, and whether Stack Bowls Three is compatible with the reported horizon-three zeros. Agreement requires documented protocol differences or corrected tables. E06E07E10E13E16
8.3 Reading coverage
Visual audit: Visually inspected the title/version/author page, all four figures, all eight tables, the method equations and Proposition 1, and Appendix A–D pages supporting training settings, baseline definitions and limitations. All six final original-PDF crops were inspected after correcting whitespace and a cut table header. Tables 1–2 retain the short caption lines defining the asterisk as reproduction by the paper’s authors. Figure 2’s routing and deployment labels were cross-checked against Eqs. (5)–(12); ambiguities are disclosed. Pages 11–14 were read as text only. Separate supplements, code and external links were not inspected; experiments were not reproduced.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 15, 16, 17, 18, 19, 20. 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; 1 Introduction (pp. 1–3)
- 2 Related Works (pp. 3–4)
- 3 Preliminaries, including 3.1–3.3 (pp. 4–5)
- 4 Methodology, including 4.1–4.2 and Proposition 1 (pp. 5–7)
- 5 Experiments, including 5.1–5.3 (pp. 7–10)
- 6 Conclusion (pp. 10–11)
- References (pp. 12–14)
- A Implementation Details, including A.1–A.3.4 (pp. 15–19)
- B Additional Experimental Results; C Limitations and future work; D LLM Usage (p. 19)
- Table 8, all per-task rows (p. 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.
- Identity scope: the inspected title page establishes arXiv:2606.05254v1, dated 3 June 2026. Title and all nine authors agree with the supplied catalog. No other revision or edition was supplied for comparison.
- Supplied acquisition omission: text extraction does not reconstruct figure images; this reading addresses that limitation by inspecting the retained PDF figures and tables.
- Separate supplemental material availability has not been fully verified; no separate supplement was supplied.
- All six supplied text chunks were read. PDF pages 1–10 and 15–20 were visually inspected; conclusion continuation and reference pages 11–14 were read as text only.
- Code, project links and external references were not opened. No experiments were run or independently reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
E01PDF p. 1, title block and arXiv margin
Title, nine authors and three affiliations match the supplied identity. The PDF is arXiv:2606.05254v1 [cs.LG], dated 3 June 2026.
Go to primary source ↓E02PDF pp. 4–5, Section 3.3, Eqs. (5)–(7); p. 5, Figure 2
Context contains past observations, actions and instruction. Video and inverse-dynamics action stages share transformer parameters and require sequential denoising passes. Figure 2 labels deployment Nv=1, Na=1/2.
Go to primary source ↓E03PDF p. 4, Sections 3.1–3.2, Eqs. (1)–(4)
Flow interpolation, shifted noise sampler and clean estimate are defined. A frozen teacher advances from σ_s to lower σ_e; student and EMA-target consistency outputs are compared. A discrete skip is introduced without a numerical setting here.
Go to primary source ↓E04PDF p. 6, Sections 4.1–4.2, Proposition 1, Eqs. (8)–(9); p. 5, Figure 2 left panel
The exact LCM velocity coefficient is −σ²σ_d/√(σ²+σ_d²); action uses b=−σ. The printed expansion gives σ²/σ_d+O(σ⁴). Proposition 1 assumes C¹ but invokes an O(σ²) Taylor remainder.
Go to primary source ↓E05PDF pp. 6–7, Section 4.2, Eqs. (9)–(12); p. 5, Figure 2 middle/right
Video retains LCM while action uses the linear endpoint map. Teacher CFG applies only to video. Weighted joint losses use a shared concatenated token sequence with flex attention. Figure 2’s middle arrows route orange action output through an f^v label; Eqs. (9)–(11) define modality maps separately.
Go to primary source ↓E06PDF p. 2, Figure 1 and Introduction; p. 3, contribution bullets; p. 7, Section 5.1; p. 9, Section 5.2
Figure 1 gives 3550+4550=8100 ms for the teacher and 193+154=347 ms for Flash-WAM 1v/1a; prose gives 348 ms. Profiling uses one L40S and an adopted 500 ms chunk budget. Page 3 distinguishes 1v/2a headline success from 1v/1a success.
Go to primary source ↓E07PDF p. 7, Table 1, teacher and 1v/2a/1v/1a rows; p. 8, Section 5.1 Benchmarks
RoboTwin has 50 tasks with Clean and Randomized splits. Teacher average is 91.25; Flash-WAM averages 85.54/81.41 with 19.0×/23.3× speedups. Teacher split values are 91.64/90.86. Naive LCM at 1v/2a has Randomized 22.07 and average 23.97.
Go to primary source ↓E08PDF p. 8, Table 2 and Section 5.1 Benchmarks; p. 9, LIBERO paragraph
LIBERO has four suites and 500 demonstrations per suite. Table 2 reports teacher 98.6, Flash-WAM 95.7/95.1 and Video-only LCM 95.2/94.2. Object-suite teacher/student 1v/2a values are 99.8/92.8. The printed Flash-WAM 1v/2a Spatial/Object/Goal/Long values are 97.0/92.8/96.4/98.0; their unweighted mean is 96.05%, differing from the printed Average of 95.7%. No weighting or other reconciliation is reported. Prose latency 6767→404 ms does not exactly yield the table’s 16.3× speedup.
Go to primary source ↓E09PDF p. 8, Figure 3 and Benchmarks paragraph; p. 9, Table 3 and Real-World Experiments
Three Unitree G1/Dex1-1 tasks use 50 demonstrations each and 10 test rollouts per task/method. T1 is pot/potato; T2 red bottle with yellow distractor; T3 pink-object target placement. Table 3 supplies the 3/10, 1/2 and 1/1 comparisons.
Go to primary source ↓E10PDF pp. 9–10, Section 5.3 and Table 4, horizon and average columns
At 1v/1a, Flash-WAM averages 82.56/80.26 versus regularized Video-only LCM 53.48/48.40. At 1v/2a, horizon-three Randomized is 63.25 versus plain Video-only LCM 68.25. Teacher averages 92.93/91.55 and naive LCM Randomized 20.08 differ from Table 1. Naive LCM horizon-three cells are zero.
Go to primary source ↓E11PDF p. 15, Appendix A.1–A.3.1 and Table 5
LIBERO teachers are fine-tuned separately for 4000 steps, then distilled for 2000. Each stage takes about 24 hours per suite on four H100s. LCM baselines share data, checkpoint and iterations; naive joint LCM changes only the action consistency parametrization.
Go to primary source ↓E12PDF p. 16, Table 6, architecture, flow, consistency and optimization rows
LIBERO configuration gives resolution 128×128, 30 action dimensions, four actions/frame, K=4, shifts 5/1, λ_a=1, EMA 0.995, σ_d=0.5, Huber c=0.001, CFG [2,10], AdamW 5e-6, batch 48 on four H100s and 100 warmup steps.
Go to primary source ↓E13PDF p. 16, Appendix A.3.2–A.3.3, Eqs. (13)–(15); pp. 7–10, Tables 1–4
Video-only LCM explicitly retains 50 action steps at inference, conflicting with table Nv/Na labels. The regularized variant adds masked action flow-matching MSE against demonstration-derived velocity.
Go to primary source ↓E14PDF pp. 17–18, Appendix A.3.4, variant overview and Eqs. (16)–(32); p. 19, Table 7/Appendix B; p. 7, Table 1
The appendix distinguishes Joint DMD2 from video-only DMD2 plus action regularization, using frozen reference, trainable student and critic. Table 1’s 1v/1a DMD2 values 52.66/48.46 match Table 7’s Joint DMD2, while DMD2 + reg. is 66.53/60.32; text assigns the regularized variant to the main table.
Go to primary source ↓E15PDF p. 19, Appendix C; pp. 8–9, Figure 3 and Table 3
Appendix C describes simulation-only experiments and physical deployment as future work, contradicting the main robot evaluation. It also identifies shared-backbone scope, missing high-noise analysis and untested multi-model/general distillation extensions.
Go to primary source ↓E16PDF p. 20, Table 8, Average row and Stack Bowls Three row; pp. 10 and 19, Tables 4 and 7
Per-task averages repeat Flash-WAM 88.42/82.66 at 1v/2a and 82.56/80.26 at 1v/1a. Stack Bowls Three is horizon 3 with naive joint LCM 24/10, conflicting with Tables 4/7’s zero horizon-three summaries.
Go to primary source ↓E17PDF p. 10, Figure 4 and caption; p. 9, Qualitative analysis
The pick_diverse_bottles visualization is an open-loop autoregressive video rollout without intermediate observation feedback. It illustrates predicted video, not executed action precision.
Go to primary source ↓8.5 Primary sources
Flash-WAM: Modality-Aware Distillation for World Action Models ↗
PDF · 10,067 extracted words
Source fingerprint
186f859f2dca5049393e47cc3207d0f66021b377516c24a00e0be6cdcb5ac678