DSWAM: A Dual-System World Action Foundation Model for Fine-Grained Robot Manipulation
1. Paper overview
In one sentence: DSWAM pairs an optional language planner with a video-co-trained executor that generates action chunks directly, trading explicit future-video inference for lower query latency. e-overviewe-architecturee-executore-objectivee-folding-protocole-sort-protocole-profile-setup
| At a glance | What to know |
|---|---|
| Research problem | Author claim Coarse household commands require semantic decomposition, while contact-rich manipulation requires precise execution. The authors argue that WAM executors lack an explicit language planning interface and that WAM–VLA comparisons often confound policy design with different robots, data and protocols. e-problem |
| Core mechanism | Source description A default WAM executor and optional vision-language planner communicate through executable language instructions; video co-training supplies supervision without requiring future-video generation at deployment. e-overviewe-architecturee-executore-objective |
| A key reported result | Matched real-world household folding: DSWAM: 96.3%, 1 min 44 s. Success rate; average completion time including timeout duration for failures. ALOHA-style dual-arm platform; shirt, skirt, pants and towel; two physical instances per category, 10 trials each. Easy and hard instances are pooled; System 2 is disabled. DeMaVLA: 92.5%, 2 min 18 s; π0: 76.3%, 2 min 26 s. Pants: DSWAM 90.0%, 2 min 19 s versus DeMaVLA 75.0%, 3 min 1 s. The rounded aggregate improvement over DeMaVLA is 3.8 percentage points and 34 seconds. Trials include retrieval, spreading and folding. This is physical executor evidence; it does not measure planner benefit or isolate video-loss causality. e-folding-protocole-folding |
| Reading caution | Reader analysis Matched platform and data reduce confounding, but no action-only-training ablation isolates video co-training. Neither the small simulation margin nor folding superiority establishes that the auxiliary video objective causes the gain. e-objectivee-folding-protocole-simulatione-folding |
Core contributions
- Source description
A default WAM executor and optional vision-language planner communicate through executable language instructions; video co-training supplies supervision without requiring future-video generation at deployment. e-overviewe-architecturee-executore-objective
- Author claim
The authors report a DeMaVLA-matched physical folding comparison, a separate sorting-supervision study, broad simulation evaluation, and TensorRT/RTC deployment diagnostics. e-folding-protocole-simulatione-sortinge-profilee-rtc
Figure 1. Language subtasks connect an optional planner to the default action executor. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the middle block: observation and task-instruction tokens enter the optional planner, whose high-level-to-low-level arrow represents decomposition. Follow the dashed yellow arrow toward the executor’s conditioning tokens. The green vision/text route supplies context, while state and noise are shown above the action expert. The right-hand backbone is explicitly labeled Wan2.2-TI2V-5B. The lower panels summarize example subtasks and execution. Read that sequence together with Section 3.2: Eqs. (4) and (8) specify the next subtask, refreshed during execution, rather than requiring an entire fixed plan to be emitted at once. The caption identifies video co-training as supervision and excludes future-video generation from inference. e-architecturee-overviewe-plannere-boundariese-executore-objectivee-folding-protocole-sort-protocol
What it supports. The diagram supports a modular interpretation: semantic instructions condition an executor whose video backbone and action expert handle continuous control. The same executor can receive the original prompt when planning is bypassed. This explains why the main folding result can evaluate WAM execution with System 2 disabled.
Where the evidence stops. The illustration separates car pickup and placement; the sorting protocol groups them into one instruction. It also omits attention-mask and freezing details. Use the written protocol and objectives for those distinctions; the graphic alone does not establish them.
2. Motivation
2.1 The problem and the proposed response
Coarse household commands require semantic decomposition, while contact-rich manipulation requires precise execution. The authors argue that WAM executors lack an explicit language planning interface and that WAM–VLA comparisons often confound policy design with different robots, data and protocols. e-problem
2.2 What this reading follows
DSWAM addresses two different bottlenecks in household manipulation: choosing an executable subtask and producing timely dual-arm actions. Its architecture assigns these jobs to separate systems, while allowing the executor to run directly from the original instruction. The crucial reading distinction is between training and deployment: future visual tokens supervise the executor during learning, but no future video is generated on the deployed action path. The experiments then separate broad simulation performance, planner-free physical folding, instruction supervision, and runtime engineering. Follow the six original visuals below to see what each comparison measures and where the paper leaves a causal or reproducibility question open. e-overviewe-architecturee-executore-objectivee-folding-protocole-sort-protocole-profile-setup
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 | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The architecture supports Dual-system: a distinct vision-language planner passes instructions to a video-based executor, and the planner can be bypassed. Other mechanisms/Outside quadrants fits action-only deployment with auxiliary video training, rather than inference-time joint future/action prediction or inverse dynamics. Efficient control is explicit; Memory & long-horizon modeling is supported more narrowly by short visual history and subtask updates, without a demonstrated persistent memory mechanism. e-overviewe-architecturee-plannere-boundariese-executore-objectivee-deployment
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 Understand what future-video supervision contributes
Begin with the current observation: three images, the active language instruction and robot state. System 1 encodes that context into latent features, then uses a conditional flow-matching action expert. During training, the same loss formulation also supervises future latent visual tokens. In Eqs. (11)–(12), the interpolation moves from target data toward Gaussian noise, and the learned velocity targets noise minus data. Equation (15) combines the action and video losses with a video weight. The key information boundary appears immediately afterward: action tokens cannot use future visual tokens as privileged input. That restriction aligns training with deployment, where only current observations are available and no future video is generated. The paper states this restriction but does not give a complete attention-mask specification or an ablation isolating its contribution. e-overviewe-executore-objectivee-reporting-gaps
5.2 Teach the planner when to advance the instruction
The planner’s supervision is more specific than labeling what a video currently depicts. A pretrained annotator segments robot trajectories into subtasks, and five-frame windows are formed from frames sampled at 1 FPS. Most windows retain their segment’s instruction. The final window of a nonterminal segment instead targets the next instruction, while the final terminal window targets done. This makes boundary detection part of language prediction. At deployment, the executor sends five recent frames every two seconds; the returned instruction stays active until another update. Equations (4) and (8) therefore describe repeated next-subtask selection. The source’s broad sequence diagram should be read through this update rule. The paper claims robustness to variable-frequency deployment input but does not quantify boundary errors or specify an automatic planner-activation criterion. e-architecturee-plannere-boundaries
5.3 Separate model quality from the timing of robot control
Reader analysis: removing future-video generation reduces one source of work, but a direct-action query can still arrive too late for smooth control. DSWAM addresses that second issue through two-engine TensorRT serving and asynchronous RTC. The visual engine builds caches; the action engine uses them for denoising updates. Meanwhile, the robot executes its current action chunk instead of waiting for each policy call to finish. Table 4 checks warmed latency and numerical action agreement; Table 5 then checks physical rollouts after changing the deployment mechanism. These are complementary measurements. Neither converts the TensorRT speedup into a guaranteed control frequency, and Table 5 combines asynchrony with RTC. A reproduction must consequently preserve both the policy checkpoint and the timing protocol to interpret the physical gain. e-executore-deploymente-profile-setupe-profilee-rtc
5.4 Training and inference
During training
Robot videos are segmented by a pretrained vision-language annotator. Five-frame sliding windows sampled at 1 FPS receive the current subtask label, the next label at a nonterminal segment’s final window, or done at the terminal final window. System 2 uses autoregressive language-model training. e-plannere-boundaries
System 1 combines action flow matching with auxiliary future-latent-video flow matching. Action tokens cannot use future visual tokens as privileged information. Figure 1 labels 5,000+ hours of real-world data; the authors assert matched pretraining and post-training data with DeMaVLA. e-architecturee-executore-objectivee-folding-protocol
RTC is applied at inference. Training predicts complete action chunks, without training-time action conditioning or a suffix-only loss. The supplied paper does not specify an optimizer schedule, frozen-module policy, video-loss weight or complete training compute. e-deploymente-reporting-gaps
During inference
Current-context features require one forward pass; no future frames are sampled, denoised or decoded. This does not mean the entire action sampler is a single pass: the deployment description includes an action-denoising engine, without specifying its iteration count. e-executore-deployment
When enabled, System 2 receives five recent frames every 2 seconds; its returned instruction remains active until the next update. Fixed 1 FPS training and variable-frequency deployment inputs are distinguished, with robustness asserted but not separately quantified. e-boundaries
BF16 TensorRT uses a visual-context engine that builds per-layer caches and an action-denoising engine consuming those caches, action latents and attention masks. Preprocessing, text encoding, VAE handling, normalization and policy wrapping remain in the serving interface. e-deployment
5.5 Implementation flow
- Choose the language condition
With System 2 disabled, use the original prompt. When enabled, a Rynnbrain4B-style planner predicts the next executable instruction from the global prompt and short visual history. The paper describes optional activation but gives no decision algorithm for when decomposition is beneficial. e-overviewe-plannere-boundaries
- Build current-context world features
Figure 1 identifies a Wan2.2-TI2V-5B pretrained video model and action expert. System 1 encodes current images, language and proprioception into latent features; its action distribution is conditioned on this representation. e-architecturee-executor
- Generate and execute a chunk
The conditional flow-matching action expert predicts continuous commands. A controller executes the current chunk while a policy worker queries the latest observations for the next one. Learned prediction and physical execution are separate processes. e-executore-deployment
6. Experiments & results
DSWAM trains a video-based robot executor with action and future-video supervision, then generates only action chunks during deployment. An optional language planner supplies subtasks. Its strongest physical comparison uses the executor alone: reported folding success is 96.3% versus DeMaVLA’s 92.5%, with timeout-inclusive completion time falling from 2 min 18 s to 1 min 44 s. The evidence supports the deployed system, while leaving the separate causal contribution of video supervision unresolved.
6.1 Read the original evidence
Table 1. Best reported simulation averages coexist with substantial task-specific weaknesses. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each method occupies two columns: Clean uses fixed initial configurations, while Rand. varies object poses and scene layouts. Compare methods within the same member of each pair before reading the average row at the bottom. DSWAM is the rightmost pair; Fast-WAM is immediately to its left. Then return to individual rows to ask whether the aggregate result hides particular failures. For example, Open Microwave and Blocks Ranking Size behave differently from the many saturated tasks. The percentages describe success in simulated bimanual tasks. Neither a large number of task rows nor high average success supplies the missing trial counts, confidence intervals or physical deployment evidence. e-sim-protocole-simulation
What it supports. DSWAM reports 92.38% clean and 91.90% randomized average success, compared with Fast-WAM’s 91.88% and 91.78%. Those are margins of 0.50 and 0.12 percentage points. They support a narrow average advantage in this table, without implying that DSWAM leads on every task.
Where the evidence stops. On randomized Open Microwave, DSWAM scores 43% while Motus scores 91%. The small average lead over Fast-WAM has no reported uncertainty estimate, so statistical reliability and broad physical generalization remain unresolved.
Table 2. The primary physical result evaluates the WAM executor with the planner disabled. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read every garment as an SR/Time pair, then compare the final Average pair. Each category pools an easier and a hard physical instance, with ten trials per instance. The procedure begins with a garment dropped into a basket and includes retrieval, spreading and the final fold. The hard instance can be absent from collection or aged and worn. System 2 is disabled throughout this benchmark, and the executor receives the original folding instruction. Section 4.1 also changes how the time column should be understood: failed trials contribute the timeout duration. Thus average completion time measures a combination of execution speed and failure frequency. e-folding-protocole-foldinge-objective
What it supports. DSWAM’s reported average is 96.3% success and 1 min 44 s, versus DeMaVLA’s 92.5% and 2 min 18 s. Pants contribute a visible improvement: 90% versus 75% success, with time reduced from 3 min 1 s to 2 min 19 s. These are physical folding outcomes under the stated matched protocol.
Where the evidence stops. The authors report matched robots, data and criteria, but this comparison changes the policy system rather than ablating the video objective. It cannot attribute the gain solely to video co-training. The supplied text does not give the numerical timeout threshold.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Matched real-world household folding ALOHA-style dual-arm platform; shirt, skirt, pants and towel; two physical instances per category, 10 trials each. Easy and hard instances are pooled; System 2 is disabled. | DSWAM: 96.3%, 1 min 44 s. Success rate; average completion time including timeout duration for failures | DeMaVLA: 92.5%, 2 min 18 s; π0: 76.3%, 2 min 26 s. Pants: DSWAM 90.0%, 2 min 19 s versus DeMaVLA 75.0%, 3 min 1 s. The rounded aggregate improvement over DeMaVLA is 3.8 percentage points and 34 seconds. Trials include retrieval, spreading and folding. This is physical executor evidence; it does not measure planner benefit or isolate video-loss causality. e-folding-protocole-folding |
| RoboTwin 2.0 bimanual manipulation 50 simulated tasks; clean fixed configurations versus randomized object poses and scene layouts. | DSWAM: 92.38 clean; 91.90 randomized. Average success rate (%) | Fast-WAM: 91.88/91.78; DeMaVLA: 88.42/86.78; Motus: 88.66/87.02; π0.5: 82.74/76.76; π0: 65.92/58.40 (clean/randomized). Best reported averages, with only 0.50/0.12 percentage-point margins over Fast-WAM. Individual tasks can regress; these simulation results do not establish physical generalization or statistical significance. e-sim-protocole-simulation |
| Real-world tabletop sorting instruction supervision Randomized object configurations; car into box and bread onto plate. Compare coarse-instruction and subtask-instruction training at 6,000 and 18,000 steps. | Subtask supervision: 100.0% and 1.00 mistakes at 6,000 steps; 100.0% and 0.30 at 18,000. Success rate; average mistakes per rollout | Raw instruction: 71.4% and 3.75 mistakes at 6,000; 80.0% and 3.30 at 18,000. Explicit subtask supervision helps this setting. Training labels change, so this is not a same-executor planner on/off test. Trial counts and the operational mistake definition are not supplied. e-sort-protocole-sorting |
| Warmed end-to-end policy inference Batch size 1; NVIDIA GeForce RTX 5090, CUDA 12.9, TensorRT 10.16.1; action agreement against PyTorch over six seeds. | BF16 TensorRT: 73.8 ms; 2.69×; 0.0106; 0.99977. Latency; speedup; maximum relative action error; cosine similarity | PyTorch: 198.2 ms, 1.00×. A measured warmed query-speed improvement with close reported action agreement, not a robot control frequency or worst-case latency guarantee. e-profile-setupe-profile |
| Asynchronous RTC on easy garments Easy instances only; synchronous TensorRT versus asynchronous TensorRT+RTC; 10 trials per task and setting. | Asynchronous: shirt 100%, 1 min 28 s; pants 100%, 1 min 8 s. Success rate; average time among successful trials | Synchronous: shirt 100%, 1 min 47 s; pants 70%, 1 min 50 s. The combined deployment change improves these small-sample rollouts. Success-only times and easy instances differ from the main folding metric; asynchrony and RTC are not separately ablated. e-profile-setupe-rtc |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 3. Explicit instruction supervision improves sorting at both reported training budgets. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Pair rows by training steps: compare the first and third rows at 6,000 steps, then the second and fourth at 18,000. This prevents extra training from being mistaken for the instruction effect. The coarse task is sorting tabletop objects; the protocol’s executable instructions place the toy car in a box and the bread on a plate. Initial object configurations are randomized. Success rate and mistakes per rollout answer different questions: the robot can ultimately finish despite intermediate mistakes. The table therefore shows why a saturated success rate still leaves room for improvement. Crucially, Section 4.1 says the executor’s training instructions change between the conditions. e-sort-protocole-sorting
What it supports. At 6,000 steps, subtask supervision gives 100% success with 1.00 mistake per rollout, versus 71.4% and 3.75 for the coarse instruction. At 18,000 steps, success remains 100% while mistakes fall to 0.30; the raw-instruction setting reaches 80% with 3.30 mistakes. The pattern favors explicit subtask context.
Where the evidence stops. This is an instruction-supervision comparison, not a fixed-executor online-planner toggle. It leaves the contribution of planner accuracy unresolved. The source supplies neither sorting trial counts nor an operational definition of an execution mistake.
Table 4. TensorRT reduces warmed policy-query latency while closely matching reference actions. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the latency and agreement columns together. The PyTorch row defines the timing baseline and action reference; dashes in its agreement columns are not zero errors measured in a separate experiment. The BF16 row reports the accelerated path. The caption specifies warmed, end-to-end policy inference at batch size one and agreement over six seeds. Section 4.5 supplies the hardware and software: NVIDIA GeForce RTX 5090, CUDA 12.9 and TensorRT 10.16.1. Section 3.3.2 explains the implementation split into visual-cache and action-denoising engines. Those details matter because this result is about a particular serving path, rather than all possible video-model deployments. e-deploymente-profile-setupe-profile
What it supports. Latency decreases from 198.2 ms to 73.8 ms, with a reported 2.69× speedup. The maximum relative error of 0.0106 and cosine similarity of 0.99977 provide an action-agreement diagnostic alongside the timing result. Together they support faster querying with close numerical agreement in the measured setup.
Where the evidence stops. Warmed batch-one timing does not specify cold-start cost, tail latency or control-loop frequency. Agreement over six seeds is a limited numerical check; it does not guarantee identical physical rollouts or establish hardware-independent performance.
Table 5. Overlapping policy queries with execution improves the reported easy-garment rollouts. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare shirt rows with shirt rows and pants rows with pants rows. Both execution settings already use TensorRT; the lower pair adds asynchronous operation and real-time chunking. The controller continues its current chunk while the policy worker computes from recent observations, as described in Section 3.3.2. Each row contains ten trials on an easy garment instance. Pay special attention to the final header, Avg. Success Time: it averages successful trials. This differs from Table 2, which combines easy and hard instances and assigns timeout duration to failures. The two tables therefore answer different questions and their absolute times should not be ranked against each other. e-deploymente-profile-setupe-rtce-folding-protocol
What it supports. Shirt success remains 100% while average success time drops from 1 min 47 s to 1 min 28 s. Pants improve from 70% to 100%, with success time falling from 1 min 50 s to 1 min 8 s. The reported benefit comes from changing deployment while retaining the trained policy.
Where the evidence stops. Asynchrony and RTC change together, so this comparison does not isolate their individual effects. Ten trials per setting and easy garments constrain the conclusion; the altered success population also affects interpretation of success-only time.
7. Analysis & limitations
7.1 What the evidence leaves open
Matched platform and data reduce confounding, but no action-only-training ablation isolates video co-training. Neither the small simulation margin nor folding superiority establishes that the auxiliary video objective causes the gain. e-objectivee-folding-protocole-simulatione-folding
The supplied experiments provide no confidence intervals, simulation trial counts or sorting trial counts. The hard folding instance may be unseen or merely aged; its category-specific status is not resolved. These are limits on robustness and generalization claims. e-folding-protocole-sim-protocole-sort-protocole-simulatione-sorting
7.2 Questions for discussion
- Does removing video co-training under matched data and compute change execution success?
- How much sorting improvement comes from training labels versus accurate online subtask transitions?
- Do asynchronous RTC gains persist on hard garments when latency and failure penalties are reported separately?
8. Reproducibility audit
8.1 Requirements and known gaps
Recover exact data partitions, checkpoints, action representation/horizon, image preprocessing, loss weight, frozen modules, optimizer settings, denoising schedule, RTC overlap rules and training hardware. The paper supplies the conceptual objectives but not these implementation settings. e-overviewe-executore-objectivee-deploymente-reporting-gaps
Resolve the planner checkpoint: Section 3.2 calls it Rynnbrain4B-style, but its citation [8] is titled Rynnvla-002. Also specify the annotator, transition-label auditing, timeout threshold, valid-fold criteria and mistake counting before attempting protocol equivalence. e-planner-referencee-plannere-folding-protocole-sort-protocol
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: Ablate video supervision while keeping the executor comparison controlled
Reader-proposed experiment, not performed: after recovering the missing training settings, initialize identical executors from the same video checkpoint and train with the full action-plus-video objective versus the action loss alone. Keep robot data, action expert, update count, seeds and deployed sampler fixed; log training compute and include a compute-matched action-only control. Disable System 2 and evaluate the same easy/hard garment instances using the full retrieval-to-fold protocol, reporting success, timeout-inclusive time and uncertainty. If video co-training is the mechanism behind the physical gain, its benefit should persist beyond seed variation and the extra-compute control. Equivalent outcomes would weaken that causal explanation without disputing the published system-level comparison. e-objectivee-executore-folding-protocole-foldinge-reporting-gaps
Check 2: Measure the gap between correct subtask labels and an online planner
Reader-proposed experiment, not performed: first reproduce Table 3’s raw-versus-subtask training comparison at matched budgets. Then freeze one subtask-trained executor and compare verified oracle subtask transitions, the learned System 2 updates, and oracle transitions deliberately delayed by one two-second update. Use paired randomized tabletop configurations and identical reset rules. Record task success, predefined execution mistakes, wrong-subtask duration and transition delay; publish trial counts and uncertainty. A large oracle-to-planner gap would show that useful supervision does not ensure reliable online decomposition. Similar degradation from the learned planner and delayed oracle would implicate transition timing. This separates training-label benefit from the deployment behavior that Table 3 leaves unresolved. e-plannere-boundariese-sort-protocole-sorting
8.3 Reading coverage
Visual audit: All 13 supplied PDF pages were rendered and visually inspected: p. 1 for title, all authors, affiliations and distinct version/date fields; pp. 2–4 for motivation, matching claims and observation/action definitions; p. 5 for Figure 1, model/data labels and planner windows; pp. 6–7 for planner boundaries, flow equations, information restrictions and deployment engines; pp. 7–8 for all evaluation protocols and metric definitions; p. 9 for every Table 1 row; p. 10 for Tables 2–3, sorting interpretation and profiling hardware/software; p. 11 for Tables 4–5 and conclusion; pp. 12–13 for references, including the unresolved planner citation. All six final original crops were separately viewed. Figure 1’s arrows were cross-checked against its caption and Sections 3.1–3.3; its sequence summary and sorting granularity were distinguished from the next-subtask update equations and two-instruction experimental protocol. The PDF contains no appendix. Separate supplements, project content, code, data and cited works remain outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13. Appendix coverage: not present.
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 page and Abstract (p. 1)
- 1 Introduction (pp. 2–3)
- 2 Related Work, including 2.1–2.3 (pp. 3–4)
- 3.1 Overview (pp. 4–5)
- 3.2 Optional System 2 High-Level Planner (pp. 5–6)
- 3.3.1 World Action Model Policy (pp. 6–7)
- 3.3.2 Real-time Chunking and Deployment (p. 7)
- 4.1 Experimental Setup (pp. 7–8)
- 4.2 RoboTwin 2.0 Simulation Benchmark (pp. 8–9)
- 4.3 Matched Real-world Folding Benchmark (pp. 8–10)
- 4.4 Optional System 2 Subtask Supervision Study (p. 10)
- 4.5 Efficiency and Deployment (pp. 10–11)
- 5 Conclusion (p. 11)
- References (pp. 12–13)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The preceding extraction limitation was addressed by inspecting all 13 PDF pages, Figure 1, all five tables, and every final crop.
- Separate supplemental material availability has not been fully verified.
- No separate supplement was supplied. The complete supplied PDF contains no appendix or dedicated limitations section.
- Code, project-page content, datasets and cited papers were not inspected; no experiments were reproduced.
- Identity/version note: the title and all 13 authors match the catalog after normalizing name order. The inspected artifact is arXiv:2607.04927v1 [cs.RO], dated 6 July 2026 in the margin, while its title-page Date is July 7, 2026. The catalog submission date is July 6. These distinct dates are preserved; no other revision was supplied or compared.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/author block, affiliation line, arXiv margin and Date field
The observed title matches the catalog. The page credits Jian Zhu, Jianjun Zhang, Taiyi Su, Tianbin Liu, Zhangyuan Wang, Kai Xie, Zitai Huang, Chong Ma, Youzhang He, Tianjian Wang, Hanyang Wang, Weihao Ding and Yi Xu; affiliations are AIRC, Midea Group and Tongji University. The margin says arXiv:2607.04927v1, 6 Jul 2026; the Date field says July 7, 2026.
Go to primary source ↓e-problemPDF p. 2, Section 1, first three paragraphs
The motivation separates semantic task decomposition from physical execution and identifies robot/data/protocol differences as confounds in WAM–VLA comparisons.
Go to primary source ↓e-overviewPDF p. 4, Section 3.1, Eqs. (1)–(2)
System 1 is the default policy; optional System 2 supplies language subtasks. Observation comprises three RGB views, language and proprioception; the output is an H-step chunk of continuous dual-arm commands. H and action-coordinate details are not numerically specified.
Go to primary source ↓e-architecturePDF p. 5, Figure 1 and caption
The overview labels 5,000+ hours of real-world data, an optional planner, vision/text encoder, pretrained Wan2.2-TI2V-5B video model and action expert. Its caption distinguishes video co-training from action-only deployment. The dashed yellow arrow passes subtask conditioning to the executor.
Go to primary source ↓e-plannerPDF p. 5, Section 3.2, Eqs. (3)–(5) and Subtask boundary supervision
The Rynnbrain4B-style planner uses five recent frames sampled at 1 Hz and a global prompt to predict the next instruction; bypass uses the original prompt. A pretrained VLM annotator segments robot videos, from which five-frame windows at 1 FPS are formed. No activation algorithm or annotator checkpoint is specified.
Go to primary source ↓e-boundariesPDF p. 6, Section 3.2, Eqs. (6)–(8) and Coupling with System 1
Final nonterminal windows target the next subtask; terminal final windows target done; other windows use the current label. Planner training is autoregressive language modeling. At deployment, five recent frames are sent every 2 seconds and the instruction is held until the next update; variable-frequency robustness is asserted.
Go to primary source ↓e-executorPDF p. 6, Section 3.3.1, Eqs. (9)–(12)
Current multimodal context produces latent features in one forward pass; no future frames are generated at inference. Actions use conditional flow matching. For actions or future visual tokens, yτ=(1−τ)y+τε and the velocity target is ε−y.
Go to primary source ↓e-objectivePDF p. 7, Eqs. (13)–(15) and following paragraph
The total loss combines action and weighted video flow matching. Action tokens are prevented from accessing future visual tokens during training. The video-loss weight is symbolic rather than numerically specified.
Go to primary source ↓e-deploymentPDF p. 7, Section 3.3.2, all paragraphs
Inference-time RTC and asynchronous querying run alongside current-chunk execution; training uses whole chunks without action conditioning or suffix-only loss. TensorRT separates visual-cache building and action-denoising engines. BF16 is used; wrapper preprocessing, encoding, VAE handling and normalization remain unchanged.
Go to primary source ↓e-folding-protocolPDF pp. 7–8, Section 4.1, Matched real-world folding protocol, Metrics and Compared methods
An ALOHA-style robot performs basket retrieval, spreading and folding for four garment categories, two instances and 10 trials each. The hard instance is unseen in collection or aged/worn. System 2 is disabled. Authors report matched robot, pretraining/post-training data and evaluation with DeMaVLA. Failure includes timeout or falling off the table; failure times use timeout duration, whose numerical value is omitted.
Go to primary source ↓e-sim-protocolPDF p. 7, Section 4.1 RoboTwin 2.0 protocol; p. 8, Section 4.2
The simulation evaluation covers 50 bimanual tasks with fixed clean configurations or randomized object poses/layouts. Per-task and average success are reported, without simulation trial counts or uncertainty estimates.
Go to primary source ↓e-simulationPDF p. 9, Table 1, Average row and Open Microwave / Blocks Ranking Size rows
Clean/randomized average SR: DSWAM 92.38/91.90, Fast-WAM 91.88/91.78, DeMaVLA 88.42/86.78, Motus 88.66/87.02, π0.5 82.74/76.76 and π0 65.92/58.40. Open Microwave is 60/43 for DSWAM versus Motus 95/91. Blocks Ranking Size is 84/85 for DSWAM versus Fast-WAM 94/98. No uncertainty is shown.
Go to primary source ↓e-foldingPDF p. 10, Table 2, all rows; p. 8, Section 4.3
Reported average SR/time: DSWAM 96.3%, 1′44″; DeMaVLA 92.5%, 2′18″; π0 76.3%, 2′26″. DSWAM shirt/skirt/pant/towel SR is 95/100/90/100%; times are 2′14″/0′58″/2′19″/1′27″. DeMaVLA pants are 75%, 3′01″. Section 4.3 reports a rounded 3.8-point SR gain and 34-second reduction over DeMaVLA.
Go to primary source ↓e-sort-protocolPDF p. 8, Section 4.1, System 2 sorting protocol and Metrics
Randomized tabletop sorting compares executors trained with a coarse command versus two subtask instructions: toy car into box and bread onto plate. Metrics are SR and average execution mistakes per rollout. Trial count, detailed mistake definition and explicit separation of online planner accuracy are not supplied.
Go to primary source ↓e-sortingPDF p. 10, Table 3, all four rows, and Section 4.4
At 6,000 steps, raw/subtask instruction SR is 71.4/100% and mistakes are 3.75/1.00. At 18,000 steps these are 80/100% and 3.30/0.30. The prose reports means across these evaluations of 75.7/100% SR and 3.53/0.65 mistakes, without trial counts or uncertainty.
Go to primary source ↓e-profile-setupPDF p. 10, Section 4.5, profiling and rollout paragraphs
Profiling uses an NVIDIA GeForce RTX 5090, CUDA 12.9 and TensorRT 10.16.1. The deployment rollout study uses easy garments and must be interpreted separately from the full easy+hard folding benchmark.
Go to primary source ↓e-profilePDF p. 11, Table 4 and caption
Warmed batch-one policy latency is 198.2 ms for PyTorch and 73.8 ms for TensorRT BF16, with reported 2.69× speedup. Action agreement over six seeds has maximum relative error 0.0106 and cosine similarity 0.99977.
Go to primary source ↓e-rtcPDF p. 11, Table 5, caption and following paragraph
Each easy-garment setting has 10 trials. Synchronous TensorRT shirt/pants SR and success-only time are 100%, 1′47″ / 70%, 1′50″; asynchronous TensorRT+RTC gives 100%, 1′28″ / 100%, 1′08″. The change combines asynchrony and RTC while retaining the trained policy.
Go to primary source ↓e-planner-referencePDF p. 5, Section 3.2 first paragraph; p. 12, Reference [8]
The planner description says Rynnbrain4B-style and cites [8], while the bibliography identifies [8] as Rynnvla-002: A unified vision-language-action and world model. The supplied PDF does not resolve the exact planner checkpoint.
Go to primary source ↓e-reporting-gapsPDF pp. 5–7, Sections 3.2–3.3; pp. 7–11, Section 4 and Tables 1–5
The source describes conceptual training objectives and deployment engines but does not supply data partitions, complete optimizer/training schedules, frozen-module policy, numerical action/video horizons, denoising steps, RTC overlap configuration or training hardware. Reported experiments do not include a video-loss-removal ablation. Inference hardware is specified separately in Section 4.5.
Go to primary source ↓8.5 Primary sources
DSWAM: A Dual-System World Action Foundation Model for Fine-Grained Robot Manipulation ↗
PDF · 6,829 extracted words
Source fingerprint
1e3360df13a89572110bfc6149bda1dffb183a50fcd0ce2b0e1888731ddef9d3