DynamicWAM: Dual-Path Motion Conditioning for World-Action Models in Dynamic Manipulation
1. Paper overview
In one sentence: DynamicWAM couples spatial flow history with image-plane motion statistics to improve interception, while cached video inference and asynchronous execution address a separate source of delay. e02e03e04e05e09e10e12e20
| At a glance | What to know |
|---|---|
| Research problem | Source description A current image can look identical when a target moves in opposite directions, yet interception requires different actions. Normalized flow helps locate motion but loses absolute image-plane displacement and elapsed time. Independently, a moving target can leave the predicted contact point while the policy computes its next chunk. e02e04e09 |
| Core mechanism | |
| A key reported result | Real-world dynamic interception and grasping: 46.67% average; linear 70.00%, circular 51.25%, compound 18.75%. Success rate. 12 tasks, four per motion level; 1,200 shared demonstrations; 20 trials/task. Success requires stable grasp/lift within 30 seconds without collision reset. π0.5: 23.75%; DynamicVLA: 21.25%; InternVLA-A1.5: 0%. The 22.92-point gain over π0.5 includes a deployment difference: DynamicWAM uses RTC; external baselines are synchronous. Compound motion remains difficult. e12 |
| Reading caution | Source description Motion comes from a fixed external camera and is not target-segmented. Distractors, occlusion and viewpoint changes can corrupt the flow/statistics. Among fifty failed physical trials, timing errors account for 42%; this selected failure sample is not a population failure rate. e15 |
Core contributions
Figure 2. Motion enters two coupled experts before future latents and actions are predicted. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the left-side arrows into two token streams. Current and rendered history-flow frames enter the frozen VAE; the future-frame arrow represents the training target that is corrupted into noisy future tokens. Proprioception, noisy actions and kinematic descriptors enter the action side. The central joint-attention block lets queries from each expert attend to keys and values from both. Text reaches video cross-attention first and influences actions through that coupling. The separate decoders produce visual predictions and robot commands. The three-stage inset distinguishes video distillation, action pretraining with the video expert frozen, and joint refinement. e03e06e07e09e12
What it supports. The architectural evidence supports a dual-system model with joint prediction: video and action experts remain distinct, yet exchange information within each layer. Kinematic tokens are conditioning inputs, not a replacement action space. During inference, video computation contributes through partial denoising and cached activations, so its role extends beyond a training-only auxiliary loss.
Where the evidence stops. The figure's “7 DOF Action” and Section 3.1's seven-dimensional state shorthand differ from Appendix A.2/Table 5's eight real-robot dimensions including the gripper. Follow the appendix for that interface. Future ground truth is a training target, not an available inference observation.
2. Motivation
2.1 The problem and the proposed response
A current image can look identical when a target moves in opposite directions, yet interception requires different actions. Normalized flow helps locate motion but loses absolute image-plane displacement and elapsed time. Independently, a moving target can leave the predicted contact point while the policy computes its next chunk. e02e04e09
2.2 What this reading follows
A robot trying to grasp a moving object needs to know both where motion occurs and how quickly the target will reach a contact point. DynamicWAM splits that evidence between a video expert and an action expert: normalized flow images describe spatial motion, while numerical tokens carry displacement and timing. Joint attention connects the two. The paper also separates learning better predictions from executing them promptly, using a compact distilled video branch, cached inference and real-time chunking. Read the simulation results as a motion-conditioning test and the physical experiments as a combined perception, prediction and execution test; their timing protocols differ. e02e03e04e05e09e10e12e20
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 | Joint prediction |
| Quadrant | Q3 · Dual-system × Joint prediction |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Dual-system × Joint prediction classification is supported by architecturally distinct video and action experts with different widths, coupled bidirectionally at each layer and trained to predict visual and action vector fields. Actions are decoded directly from their own stream; this is not inverse dynamics applied after completed video generation. Video computation persists during inference through partial denoising and cached activations. e03e06e09
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 Recover the information that normalized motion images discard
Begin with an observed interval rather than a predicted future. The model computes displacement between two timestamped head-camera frames and rejects unreliable flow vectors. Rendering the remaining field makes direction and spatial extent visible to a pretrained video encoder. However, per-frame percentile normalization means a larger displacement can produce the same image, and no pixel color records the interval's elapsed time. DynamicWAM therefore computes statistics before rendering: signed mean displacement, mean and percentile magnitude, duration, velocity and acceleration. These become four learned kinematic tokens, one per history interval. Their units remain image-plane pixels and physical seconds. Learned invalid embeddings distinguish missing history from valid small motion. This explains the representation's intended complementarity without assuming that the statistics identify a particular object or recover three-dimensional dynamics. e04e05e08
Figure 3. One observed flow field supplies spatial appearance and numerical motion evidence. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the upper row from the two endpoint images through Farnebäck estimation and the consistency filter to the RGB rendering. Its direction-to-hue and magnitude-to-intensity labels describe a normalized representation: Equation (4) divides magnitude by the interval's own 99th percentile. The lower row starts from the same masked field, retaining numerical displacement statistics and adding timestamp duration. Velocity divides displacement by duration; acceleration compares neighboring interval velocities. Standardization and an MLP map these twelve numbers into tokens. The main text additionally specifies position/type embeddings and learned handling for invalid intervals or unavailable acceleration predecessors. e04e05e08e15
What it supports. The two outputs retain different information. Rescaling a flow field by a positive factor can leave its normalized picture unchanged, while the numerical descriptor changes. Duration is also absent from the picture. This construction explains why a separate token path could help estimate contact timing even when flow images already reveal the motion direction.
Where the evidence stops. These statistics use flow-grid pixels and seconds, not calibrated metric-space target velocity. They summarize valid image regions without isolating the instructed object. Robot motion, distractors or occlusion can therefore contaminate both paths despite the consistency filter.
5.2 Follow joint prediction from supervision to a robot command
The architecture connects a compact video transformer to a narrower action transformer through joint attention. It does not first complete a video and then invoke an inverse-dynamics module. During training, clean future visual latents and clean actions are separately mixed with noise; both experts predict flow-matching vector fields. Video distillation precedes action pretraining, followed by joint refinement with a small visual-loss weight. During inference, future observations are unavailable: the model denoises predicted future latents and action tokens using current observations, history, language and state. To reduce computation, full video forwards run only during the first two of ten action denoising steps; later steps reuse cached video activations. The action decoder outputs the robot's existing command representation. Executing those commands and refreshing observations close the physical feedback loop. e03e06e07e09
5.3 Separate representation quality from the timing of execution
Reader analysis: the cleanest representation comparison is the synchronous DOMINO ablation, because it holds the backbone and training schedule fixed while changing available motion information. Its black-flow control also preserves temporal input geometry. The physical-robot comparison answers a broader question: how well does the whole deployed system intercept a moving target? External baselines and DynamicWAM use different execution protocols there. Table 11 narrows that question by keeping full-model weights fixed and separating standard synchronous operation, RealCore-only execution and RTC. This distinction matters because DOMINO pauses while inference runs, whereas a real target keeps moving. The appearance-shift results add another boundary: DynamicWAM starts from higher in-distribution success and keeps an absolute advantage, but its retention ratio is similar to the comparators'. Better absolute OOD performance therefore does not demonstrate improved proportional robustness. e11e12e13e14e20
5.4 Training and inference
During training
Stage 1 distills Wan2.2-TI2V-5B using ground-truth flow matching, PCA-projected hidden states and temporal-difference matching. Stage 2 freezes the compact video expert and trains the action/kinematic modules. Stage 3 jointly refines both experts while freezing the VAE and text encoder. e07
DOMINO stages run 80k/80k/40k steps; real-robot schedules are 8k/15k/8k. Stage 2 initializes the action expert and fixed action-normalization statistics from a pretrained history-flow-only checkpoint. During joint refinement, the video-loss weight decreases from 0.01 to 0.001 over 2,000 steps. e07
During inference
Ten action denoising steps use full video forwards only at steps 0 and 1, then reuse cached video keys/values. The appendix reports approximately 550 ms for full joint denoising versus 173.7 ms with caching. This is computation within the predictive model, not physical action execution. e09
DOMINO executes all sixteen actions synchronously and pauses simulation during inference. Real-robot RTC prefetches when fewer than four queued steps remain and merges chunks over four steps with maximum guidance weight 0.5. It overlaps computation and execution; it does not shorten an individual query. e09e10e20
5.5 Implementation flow
- Build aligned motion history
Four intervals use endpoints spaced four policy steps apart. Farnebäck flow on 64×64 head-camera grayscale images is filtered by forward–backward consistency. Direction becomes hue; magnitude is divided by its valid-pixel 99th percentile. The resulting images and current observation are encoded temporally through the frozen VAE. e04e08
- Restore magnitude and elapsed time
Each interval supplies signed displacement, mean and 99th-percentile magnitude, duration, corresponding velocity statistics, and acceleration. Dataset-standardized descriptors pass through a two-layer MLP plus position/type embeddings. Invalid intervals and missing acceleration predecessors use learned embeddings. These are image-plane quantities, not recovered 3-D object velocities. e05
- Couple distinct experts
Both video and action queries attend to concatenated keys/values from both streams. Text enters through video cross-attention. The action stream has one state, sixteen action, four kinematic and four register tokens; decoding reads the state/action positions. The 12-layer experts have widths 2048 and 768, with 988.8M trainable parameters during joint refinement. e03
6. Experiments & results
DynamicWAM gives a coupled video/action model two views of recent motion: normalized optical-flow images preserve spatial structure, while numerical tokens restore displacement scale and timing. A distilled video expert and cached inference support deployment with asynchronous action chunks. Reported success is 38.2% on synchronous DOMINO and 46.67% across twelve physical-robot tasks, subject to the stated protocol and robustness boundaries (e03, e04, e05, e09, e10, e12, e14).
6.1 Read the original evidence
Table 1. DynamicWAM leads the reported task metrics, with a measurable query-time cost. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row across three separate quantities: lower query latency is desirable, while higher success rate and manipulation score are desirable. MS records partial benchmark progress rather than binary completion. The common setup is thirty-five clean dynamic Level-1 tasks, one hundred accepted evaluation episodes per task with unseen instructions, and matched fine-tuning data of 10,500 demonstrations. Compare the final row with InternVLA-A1.5, the strongest listed competitor on SR and MS. Appendix A.4 measures latency over one thousand queries on one RTX 5090, including preprocessing and motion construction. e10e18e20
What it supports. DynamicWAM reports 38.2% success and 53.2 MS, exceeding InternVLA-A1.5 by 8.9 percentage points and 10.7 MS points. Its 173.7 ms query takes longer than that baseline's 158.4 ms. The table therefore supports an accuracy advantage under the stated protocol, without supporting superiority on every efficiency measure.
Where the evidence stops. Section 4.2's claim of a 3.2× speedup over InternVLA-A1.5 conflicts with these latency cells. The appendix's approximately 550-to-173.7 ms comparison concerns DynamicWAM caching instead. DOMINO also pauses during inference, so the table does not test target displacement during computation.
Table 2. Physical interception improves across motion levels, while compound trajectories remain challenging. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with L1, L2 and L3: these denote linear, circular and compound motion, each containing four physical-robot tasks. The average column weights the twelve tasks equally. Every policy is fine-tuned on the same 1,200 demonstrations and evaluated for twenty trials per task. Appendix A.4 counts success only when the robot stably grasps and lifts the target within thirty seconds without a collision reset. Compare external baselines separately from the motion-path variants. The full DynamicWAM entry includes RTC asynchronous execution; Section 5.2 says external baselines follow their official synchronous deployment protocols. e12e13
What it supports. The full model achieves 46.67% average success versus 23.75% for π0.5, a 22.92-percentage-point difference. Its rates are 70.00% on linear motion, 51.25% on circular motion and 18.75% on compound motion. The latter result exceeds the zero successes recorded by external baselines but still leaves most compound trials unsuccessful.
Where the evidence stops. The external comparison includes execution differences, so it cannot attribute the entire gain to motion conditioning. Twenty trials per task give limited precision, and confidence intervals are not reported. Table 11 provides the more controlled comparison for the execution mechanism.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| DOMINO Level 1 dynamic manipulation 35 tasks, clean dynamic setting, unseen instructions; 100 accepted episodes/task; 10,500 training demonstrations; native synchronous execute-16 protocol. | 38.2% SR; 53.2 MS; 173.7 ms. Success rate (SR), manipulation score (MS), per-query latency | InternVLA-A1.5: 29.3% SR, 42.5 MS, 158.4 ms; gains are 8.9 percentage points and 10.7 MS points. Best reported SR/MS, with slower queries than this baseline. MS measures partial task progress. The prose claim of a 3.2× speedup over InternVLA-A1.5 contradicts Table 1. e10e18 |
| Dual-path motion ablation on DOMINO Same backbone/training schedule and synchronous evaluation; the no-flow control uses black frames to preserve input geometry. | Full: 38.2% / 53.2. SR / MS | Neither path: 22.7% / 38.3; flow only: 27.2% / 41.6; tokens only: 30.5% / 47.2. Both paths improve the aggregate; full gains 7.7 SR points over tokens alone. Benefits vary: Put Object Cabinet falls from 17% with flow only to 7% with both. e11 |
| Real-world dynamic interception and grasping 12 tasks, four per motion level; 1,200 shared demonstrations; 20 trials/task. Success requires stable grasp/lift within 30 seconds without collision reset. | 46.67% average; linear 70.00%, circular 51.25%, compound 18.75%. Success rate | π0.5: 23.75%; DynamicVLA: 21.25%; InternVLA-A1.5: 0%. The 22.92-point gain over π0.5 includes a deployment difference: DynamicWAM uses RTC; external baselines are synchronous. Compound motion remains difficult. e12 |
| Real-world execution-component ablation Identical full-model weights, histories, horizons and inference settings across the twelve-task suite. | RealCore plus RTC: 46.67%. Average success rate | Standard synchronous: 37.50%; RealCore only: 42.08%. RealCore adds 4.58 points; RTC adds another 4.59. Table 3's 42.08% no-RTC comparator already includes RealCore. e13 |
| Level 1 appearance-shift generalization Four linear-motion tasks; 20 trials/task/condition; background, lighting and object-color shifts preserve motion and task semantics. | 36.3% OOD SR; 0.52 retention. OOD success rate; OOD/ID retention | π0.5: 17.5%, 0.54; current-frame WAM: 20.0%, 0.53. Higher absolute shifted success does not establish stronger proportional robustness. Background-shift success is only 25.0%. e14 |
| Stage-1 future-frame reconstruction 500 held-out DOMINO packed windows; ten flow-matching steps; frozen-VAE decoding with a shared first-latent protocol at 192×160. | 0.291; 0.0565; 25.0 dB. Latent RMSE; pixel RMSE; PSNR | Ground-truth future latents/decoded frames; no comparative baseline metric is reported here. This assesses the distilled video expert's reconstruction, not closed-loop robot success or the causal value of imagination. e16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 4. The combined motion representation improves both aggregate task metrics. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Decode the removal labels before comparing heights. “w/o flow and motion” removes both motion paths. “w/o motion” retains rendered flow but removes kinematic tokens. “w/o flow” retains the tokens while substituting constant black flow frames, preserving temporal input geometry. The hatched final bar uses both paths. The left panel reports percentage success and the right reports manipulation score; their identical vertical ranges do not make the units interchangeable. Appendix A.5 states that these variants share the backbone and training schedule, and the DOMINO experiment keeps RTC disabled. e11
What it supports. Flow alone raises SR from 22.7% to 27.2%; tokens alone reach 30.5%; together they reach 38.2%. The corresponding MS values follow the same ordering. The full model gains 7.7 SR points and 6.0 MS points over the stronger single-path variant, supporting useful complementary information in the reported aggregate.
Where the evidence stops. The chart provides no uncertainty bars or repeated-seed variation. Aggregate improvement is not universal task improvement: Table 8 reports 17% success for flow only but 7% for the full model on Put Object Cabinet. The proposed mechanism needs task-level scrutiny.
Table 11. Execution infrastructure and asynchronous chunking make separate contributions. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the two binary columns as successive changes to deployment. The first row disables both RealCore and RTC. The second enables the real-time execution core while retaining synchronous chunking. The third adds RTC to that core. Appendix D.4 states that model weights, observation histories, action horizons and inference settings are held fixed. This table clarifies the main paper's Table 3: its 42.08% “w/o RTC” row is the RealCore-only setting, not the standard synchronous row shown here. RTC generates a new chunk while queued actions are still executing. e09e13
What it supports. RealCore raises average success from 37.50% to 42.08%, a 4.58-point gain. RTC then raises it to 46.67%, another 4.59 points; the total improvement over standard synchronous execution is 9.17 points. These are changes in executed-task success, not measurements of faster individual model queries.
Where the evidence stops. The three rows do not include RTC without RealCore, so they are not a complete factorial experiment. RealCore is not specified as a standalone implementation in this PDF. The result supports the reported sequence of execution changes; it does not isolate every possible interaction.
7. Analysis & limitations
7.1 What the evidence leaves open
Motion comes from a fixed external camera and is not target-segmented. Distractors, occlusion and viewpoint changes can corrupt the flow/statistics. Among fifty failed physical trials, timing errors account for 42%; this selected failure sample is not a population failure rate. e15
Reported tables do not provide confidence intervals or repeated-training-seed variation. Attention overlays are coarse, averaged diagnostic maps; they do not establish target-specific causal grounding. e10e12e17
Figure 2 and Section 3.1 describe the real arm as seven-dimensional, whereas Appendix A.2 and Table 5 specify eight state/action dimensions including the gripper. This report follows the explicit implementation table and preserves the discrepancy. e03e12
7.2 Questions for discussion
- Do kinematic tokens still help when flow is rendered with an externally calibrated magnitude scale?
- How much advantage remains when all physical-robot baselines receive the same asynchronous execution infrastructure?
8. Reproducibility audit
8.1 Requirements and known gaps
The recipe uses eight NVIDIA H100 GPUs, global batch 128, AdamW, bfloat16 and DeepSpeed ZeRO Stage 0. Latency uses 1,000 queries on one RTX 5090 and includes preprocessing and motion construction. Reproducing timings requires the stated video-cache schedule. e07e10e09
Required artifacts include the teacher, frozen encoders, the history-flow-only initialization checkpoint, action-normalization statistics and demonstrations. Exact checkpoint identifiers, a complete standalone RealCore implementation, software versions and wall-clock training cost remain unspecified. Preserve actual timestamps and validity masks when reconstructing the supplied flow recipe. e07e08e13
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: Test whether the kinematic path actually distinguishes speed and duration
Reader-proposed check, not performed: use paired simulated interception trajectories with the same direction and similar normalized flow patterns but different displacement magnitudes or timestamp durations. Hold the current view, task, motion-history length and synchronous execution protocol fixed as closely as the environment permits. Compare full conditioning with flow only, then intervene on the full model by shuffling kinematic descriptors between paired histories while preserving validity masks. Record contact-time error and success by speed, alongside rendering similarity. If correctly paired descriptors improve timing but shuffled descriptors remove that gain, the claimed scale/timing mechanism receives more specific support. No selective effect would weaken that interpretation. e04e05e08e11
Check 2: Measure execution gains under matched target motion and measured observation age
Reader-proposed check, not performed: use one fixed full-model checkpoint and replay matched target-trajectory conditions across standard synchronous, RealCore-only and RealCore-plus-RTC deployment. Match camera streams, action horizon, denoising/cache schedule and success criterion; randomize execution order and repeat trials. Log observation capture, query completion, action application, queue starvation and contact-time error, in addition to success. Repeat at several imposed inference delays without changing model outputs' representation. RTC should reduce pauses and observation-to-execution delay without reducing isolated query duration. A success gain unaccompanied by improved timing, or one explained by different observation histories, would challenge the proposed execution explanation. e09e12e13e20
8.3 Reading coverage
Visual audit: All 18 original PDF pages were rendered at 200 DPI and visually inspected, including the title/version block, complete method and equations, evaluation protocols, training/hardware details, appendices, all nine figures and all twelve tables. Six original crops were inspected individually: Figures 2–3, Table 1, Figure 4, Table 2 and Table 11. Narrow visuals were cropped from 400-DPI renders; the architecture uses 200 DPI. Their arrows, normalization labels, token paths, table units and removal labels were cross-checked against the text. The architecture's action-dimension discrepancy and Table 1's conflicting latency prose are disclosed. Figures 8–9 and their diagnostic protocols were inspected but not selected as crops. Separate supplements, code and external resources remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 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
- Abstract
- 1 Introduction
- 2 Related Work
- 3 Method, including 3.1–3.4
- 4 Simulation Experiments, including 4.1–4.4
- 5 Real-World Experiments, including 5.1–5.5
- 6 Conclusion
- References
- A Experimental Details, A.1–A.5
- B Architecture and Training Details, B.1–B.6 (B.2 is titled Optical-Flow Estimation but contains flow-matching parameterization)
- C Motion Representation Construction, C.1–C.4, including C.2 Forward–Backward Consistency Filtering
- D Extended Experimental Results, D.1–D.8
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 complete supplied 18-page PDF was read through all six text chunks; all 18 pages and all nine figures and twelve tables were visually inspected. The acquisition warning that text extraction does not reconstruct figure images was addressed by inspecting the PDF renders.
- Separate supplemental material availability has not been fully verified; no separate supplement was supplied.
- The inspected title page identifies arXiv:2608.00793v2 [cs.RO], 6 August 2026. Its nine authors match the catalog. The title uses World–Action typographically; the supplied observedTitle uses World-Action. No substantive title or author difference was found. Version 1 and revision history were not supplied, so changes between versions remain unverified.
- Code, project pages, external references and checkpoints were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block and arXiv margin
DynamicWAM title, nine named authors, five affiliations, and arXiv:2608.00793v2 [cs.RO], 6 August 2026, establish the supplied edition's identity.
Go to primary source ↓e02PDF pp. 1–2, Introduction and Figure 1
Similar current images can require different contact actions; history flow and numerical motion descriptors address direction, scale and timing ambiguity.
Go to primary source ↓e03PDF p. 3, Figure 2 and Section 3.1, Eqs. (1)–(2); p. 11, B.1 and Table 5
Distinct video/action experts use joint attention. Table 5 reports 12 layers per expert, widths 2048/768, 988.8M trainable parameters, sixteen actions and eight future RGB frames. B.1 specifies token ordering and video-only text cross-attention. Main-text seven-dimensional real-state notation differs from Table 5's eight dimensions including gripper.
Go to primary source ↓e04PDF pp. 3–4, Section 3.2, Figure 3 and Eqs. (3)–(7)
Four history intervals with stride four produce masked flow renderings, encoded with the observation through the frozen VAE. Per-frame percentile normalization removes global displacement scale and omits duration.
Go to primary source ↓e05PDF p. 4, Section 3.2, Eqs. (8)–(13); p. 13, C.4 and Table 7
Twelve-dimensional descriptors combine displacement, duration, velocity and acceleration statistics, then use standardization and learned token/invalid-state embeddings. Table 7 units are flow-grid pixels and simulator seconds.
Go to primary source ↓e06PDF p. 4, Section 3.3, Eqs. (14)–(15)
Visual and action streams regress conditional-flow-matching vector fields with a weighted sum of squared-error losses.
Go to primary source ↓e07PDF pp. 11–12, B.3–B.5 and Table 6
Distillation uses ground-truth, hidden-state and motion losses. Staged training, history-flow-only checkpoint initialization, inherited normalization, freezing, schedules, video-loss annealing and eight-H100 optimization settings are specified; exact artifact identifiers and wall-clock training duration are not given.
Go to primary source ↓e08PDF pp. 12–13, C.1–C.4 and Eq. (16)
The recipe specifies raw-frame offsets 48/36/24/12/0, timestamp-based derivatives, 64×64 Farnebäck estimation and consistency filtering. Thresholds are 0.01 relative, 0.5 absolute and 0.1 minimum accepted fraction.
Go to primary source ↓e09PDF p. 5, Section 3.4; p. 12, B.6
RTC overlaps inference/execution with a four-step prefetch threshold and overlap, maximum guidance 0.5. Ten action denoising steps use full video forwards at 0/1 then cached video activations; latency is approximately 550 ms without caching versus 173.7 ms with it.
Go to primary source ↓e10PDF p. 5, Sections 4.1–4.2 and Table 1, DynamicWAM and InternVLA-A1.5 rows; p. 10, A.1, A.4–A.5
DOMINO uses 35 Level-1 clean dynamic tasks, coefficient 0.1, unseen instructions and 100 accepted episodes/task after training on 10,500 demonstrations. Table 1 reports 38.2/53.2/173.7 versus 29.3/42.5/158.4 for SR/MS/latency. MS measures partial progress; query latency includes preprocessing on one RTX 5090 over 1,000 queries.
Go to primary source ↓e11PDF p. 5, Figure 4 and Section 4.3; p. 11, A.5 variant definitions; p. 14, Table 8, Overall and Put Object Cabinet rows
Neither/flow-only/tokens-only/full variants report SR 22.7/27.2/30.5/38.2 and MS 38.3/41.6/47.2/53.2. The no-flow variant retains black frames. Put Object Cabinet is 17% with flow only and 7% full.
Go to primary source ↓e12PDF p. 6, Sections 5.1–5.3 and Table 2; p. 10, A.2–A.4; p. 15, Table 9
Twelve physical tasks use 1,200 demonstrations and twenty trials/task on FR3 arms with two D435i cameras and stated grippers. Full RTC averages 46.67%, versus π0.5 23.75%, with level rates 70.00/51.25/18.75. External baselines execute synchronously; success is stable grasp/lift within thirty seconds without collision reset.
Go to primary source ↓e13PDF p. 7, Section 5.4 and Table 3; p. 13, D.4; p. 16, Table 11
Identical full-model configurations yield 37.50% standard synchronous, 42.08% RealCore-only, and 46.67% RealCore plus RTC; the latter two correspond to Table 3's comparison.
Go to primary source ↓e14PDF p. 7, Section 5.5 and Table 4; p. 10, A.4; p. 13, D.3; p. 15, Table 10
Four Level-1 tasks receive three appearance shifts with twenty trials/task/condition while other task settings are fixed. DynamicWAM OOD SR is 36.3%, retention 0.52, and background SR 25%; π0.5 is 17.5%/0.54 and current-frame WAM 20.0%/0.53.
Go to primary source ↓e15PDF p. 7, Section 6; p. 16, D.8
Authors identify fixed-camera 2-D motion, lack of instructed-target isolation, distractors, occlusions and viewpoint sensitivity. Fifty failed physical trials are categorized, with 42% timing errors.
Go to primary source ↓e16PDF p. 13, D.5, continued on p. 16; p. 17, Figure 8
Stage-1 evaluation uses 500 held-out packed windows and ten sampling steps with shared first-latent decoding. Latent RMSE is 0.291, pixel RMSE 0.0565 and PSNR 25.0 dB; Figure 8 contrasts GT/predicted futures for three tasks.
Go to primary source ↓e17PDF p. 16, D.6–D.7; p. 18, Figure 9
Kinematic-to-condition attention is recomputed from projected queries/keys, averaged across heads/tokens and layers 8/10/11 at noise level 0.15, then projected from the coarse latent grid. Additional text ablations vary history length, stride and rendering percentile.
Go to primary source ↓e18PDF p. 5, Section 4.2 latency paragraph and Table 1; p. 12, B.6 inference acceleration
Section 4.2 claims approximately 3.2× faster queries than InternVLA-A1.5, but Table 1 lists 173.7 ms versus 158.4 ms. B.6 separately reports approximately 550-to-173.7 ms for DynamicWAM's own caching comparison; this does not validate the baseline speedup claim.
Go to primary source ↓e20PDF p. 10, A.4 DOMINO protocol; p. 11, A.5 Sim2Real Performance Gap
DOMINO executes sixteen steps at 100 ms intervals and suspends simulation during policy inference. Physical targets continue moving during inference, limiting transfer of latency conclusions between protocols.
Go to primary source ↓8.5 Primary sources
DynamicWAM: Dual-Path Motion Conditioning for World-Action Models in Dynamic Manipulation ↗
PDF · 8,765 extracted words
Source fingerprint
39833e174e8732fd18c25c4a0b287fbc00b7c0567d5bdab88eda290173cd2acf