DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation
1. Paper overview
In one sentence: Separating base, arm, and ego-motion information improves FastWAM's replay prediction and observed robot coordination with few Stage-2 trainable parameters, while retaining a large, slightly slower deployed model. e02e03e04e05e11e12e14e15e16
| At a glance | What to know |
|---|---|
| Research problem | Source description A body-mounted camera mixes scene motion with base-induced viewpoint changes. Meanwhile, arm/gripper targets and base velocities have different semantics and control rates. DECOWAM gives these factors explicit conditioning paths while retaining a joint video–action prediction interface. e02e03 |
| Core mechanism | |
| A key reported result | Real-robot whole-body coordination and robustness: DECOWAM: 44.3% / 30.4% / 32.9%. Coordination / base-displacement robustness / autonomous recovery success ↑. Table V(b), 79 trials per method; reported percentages rounded to one decimal. FastWAM: 34.2% / 12.7% / 27.8%. Observed coordination and displacement-robustness gains are 10.1 and 17.7 percentage points. Operational thresholds and perturbation magnitudes are not specified. e15 |
| Reading caution | Reader analysis The dataset bookkeeping is unresolved: Section V describes 795 raw HDF5 episodes before reporting 1,487 filtered episodes, without explaining that transition. The 217/214-episode converted subsets and replay slice also require immutable membership manifests to audit separation. e09 |
Core contributions
Figure 1. Four compact adaptation paths specialize an aligned video–action backbone for whole-body control. Original paper, p. 2 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with panel C: pretrained FastWAM is aligned to ARMDOG, then the decoupled modules are adapted. Panel A expands those modules around the backbone. Green adapters correct intermediate WAN features; the pink box transfers future information into a student; the green dual-latent box separates base and arm information; the yellow path projects current base velocity into an ego-motion condition. Follow the arrow legend to distinguish conditioning from training signals. The output boxes specify eight future frames and a 48-step, 14-D action chunk. Finally, panel B crosses out future frames: the deployed policy receives instruction, RGB, and current state. e03e04e05e06e07e08e16
What it supports. The design introduces explicit routes for three entangled quantities: what the base does, what the arm does, and how base motion changes the camera view. Freezing the aligned backbone confines Stage-2 updates to added modules, while the current-input deployment path retains learned future-informed conditioning.
Where the evidence stops. This is a schematic: Section IV distinguishes WAN video flow from ActionDiT action flow even though the figure groups output arrows beneath ActionDiT. The deployable-policy block does not specify an execution cadence or demonstrate low inference latency.
2. Motivation
2.1 The problem and the proposed response
A body-mounted camera mixes scene motion with base-induced viewpoint changes. Meanwhile, arm/gripper targets and base velocities have different semantics and control rates. DECOWAM gives these factors explicit conditioning paths while retaining a joint video–action prediction interface. e02e03
2.2 What this reading follows
A mobile manipulator must distinguish an object moving from its own camera moving, while coordinating an arm with a locomoting base. DECOWAM builds this distinction into an existing video–action model through dedicated conditioning paths. Its future-aware student learns from a teacher that can inspect later frames, but deployment uses only current observations, state, and language. The evidence is strongest when read in layers: matched replay improves both output branches; internal ablations support the added paths as a package; physical trials show larger coordination and displacement-robustness gains than final-success gains. Parameter efficiency describes the final adaptation stage, not the size or speed of the deployed network. e02e03e04e05e11e12e14e15e16
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
Conflicts with the recorded classification
WAM and Dual-system are supported by the separate WAN video and ActionDiT action experts. However, Eq. (1) and deployment explicitly describe joint future-video/action prediction, which conflicts with labeling the prediction paradigm only Other mechanisms and placing it Outside quadrants. There is no described video-to-action inverse-dynamics pipeline. This is a reader assessment; the recorded catalog snapshot is preserved. e03e08
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 Learn from the future, then remove access to it
Start with the distinction between a prediction target and an input available to the robot. DECOWAM predicts future images and actions, but those future images are also privileged teacher inputs during training. Their WAN-VAE latent summaries help the teacher organize examples by action-related outcomes. The student receives only current-image statistics and state, reconstructs actions, and matches the teacher through a stopped-gradient target. Only this student representation conditions the action expert. At deployment, the teacher and auxiliary heads disappear, while the student and both generative experts remain. Stage 2 also freezes the already aligned FastWAM weights; it does not erase the cost of Stage 1 or compress the full network. Parameter-efficient specialization and causal deployment are therefore separate properties. e04e05e08e16
Figure 2. Privileged future observations shape a student representation without becoming deployment inputs. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the blue paths as information available now: current RGB becomes a WAN-VAE summary c, and current state s_0 enters both encoders. The red dotted path supplies future summary f only to teacher q_t during training. The student q_s has no corresponding future input. On the right, purple blocks show action reconstruction, stopped-gradient teacher–student matching, and geometry supervision that relates latent similarity to action similarity. Equation (9) also reconstructs actions from the student. At the bottom, only the student's representation passes through bias head B_q into the action context, connecting representation training to deployable action prediction. e05e08e12
What it supports. The transfer happens through supervised representation shaping. The teacher sees information unavailable at decision time, while the student learns to approximate action-relevant structure from present inputs. Thus, future frames support training without requiring a deployed teacher or access to the robot's eventual outcome.
Where the evidence stops. The label quotient does not prove that latent equivalence captures physical outcome equivalence. Table II removes the entire bottleneck; it does not isolate geometry supervision from distillation and action reconstruction.
5.2 Give base velocity two distinct jobs
The same three base-velocity channels have different responsibilities in the two prediction branches. In the action branch, base velocity is a future target, distinct from the arm and gripper targets. Separate 16-D factors represent these control subsets. Direct heads train each factor to preserve its assigned information, while cross heads try to predict the opposite subset. Gradient reversal lets those heads learn their prediction task but pushes the factor encoders against cross-task predictability. Their concatenated factors then bias action context. In the video branch, current normalized base velocity is an explanatory condition for viewpoint motion, projected into every token. This is learned conditioning, not camera calibration or geometric image warping, and the training objective encourages rather than guarantees independent factors. e03e06e07
5.3 Keep prediction, mechanism, and execution evidence separate
Reader analysis: the strongest conclusion depends on keeping three comparisons distinct. Table I establishes a replay improvement from a particular FastWAM initialization. Table II supports added conditioning paths within another training suite, but does not isolate gradient reversal or the future bottleneck's individual losses. The physical tables then establish observed task and coordination outcomes under the robot protocol. They do not demonstrate that generating video itself caused those gains, because the complete architecture changes several mechanisms at once. X-VLA's stronger action-only replay errors further prevent treating DECOWAM as the best action predictor. A useful research hypothesis is that explicit embodiment factors help maintain coordination under moving viewpoints; the larger displacement and transport margins motivate that hypothesis, while comparable final completion and missing uncertainty constrain it. e11e12e13e14e15
5.4 Training and inference
During training
Both experts use conditional flow matching between Gaussian noise and their targets. The Stage-2 objective adds bottleneck and disentanglement losses with weights 0.2 and 0.1 to unit-weight video/action losses. Both loss gates and action-context injection gates remain one in the reported run. e08
The converted snapshot has 217 episodes, 27 task folders, and 56,041 frames; Stage 2 uses 214 episodes from 26 tasks after excluding Legacy val. The larger quality-filtered corpus is separately reported as 1,487 episodes and 343,550 frames. These are distinct counts, not interchangeable training-set sizes. e09
During inference
Remove the teacher and auxiliary prediction heads. Compute the student, base/arm factors, and current-velocity condition from current inputs, then sample both flows. The paper reports closed-loop physical execution with common low-level control and safety constraints, but does not specify a replanning cadence or an explicit planner that scores generated video. e08e14
5.5 Implementation flow
- Preserve the two-expert interface
A Wan-2.2 video expert predicts flow over future visual latents, while ActionDiT predicts flow over action chunks. Both receive language and proprioceptive context. Joint prediction does not imply a single shared transformer. e03e10
- Adapt after alignment
Stage 1 adapts FastWAM to ARMDOG for 50k steps. Stage 2 freezes that checkpoint and trains the four added parameter groups. WAN-block residual adapters use a 128-D down-projection, SiLU, and an up-projection. e04
- Transfer privileged future information
WAN-VAE latent means and standard deviations summarize current and future images. A 64-D teacher sees both summaries and state; the student sees only the current summary and state. Action reconstruction, stopped-gradient distillation, and batch-median-scaled action/teacher-distance matching supervise this bottleneck. Only the student biases action context. e05e10
- Separate control factors and expose ego-motion
Pooled action context produces separate 16-D base and arm latents. Direct heads reconstruct each assigned action subset; gradient-reversal cross heads oppose encoding the other subset. Their concatenation biases action context. Separately, projected current base velocity biases every video token; it supplies conditioning rather than explicit geometric warping. e06e07
6. Experiments & results
DECOWAM adapts FastWAM to a moving quadruped–arm platform by separating arm control, base control, and camera ego-motion. It jointly predicts future video and actions using small trainable additions to an already adapted backbone. Replay errors improve, while hardware evidence favors coordination and displacement robustness more clearly than final task completion.
6.1 Read the original evidence
Table I. The closest replay comparison improves both visual and action prediction from the strongest observed FastWAM checkpoint. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First locate the Original FastWAM 50k row: it is the initialization used for decoupled adaptation. The 40k and 80k rows show why the stopping point matters; neither is a stronger baseline here. Compare the final row with 50k across all three columns. Lower frame MSE measures pixel error, higher PSNR measures pixel fidelity in decibels, and lower action MSE measures error in normalized 14-D action space. Section VI.A fixes current inputs, normalization, and 16 replay batches from the 23-episode Box-val slice. Each FastWAM-family prediction contains eight 384 × 320 frames and a 48-step action chunk. e03e10e11
What it supports. DECOWAM reports frame MSE 8.77e-4 and action MSE 5.38e-5, versus 1.032e-3 and 6.87e-5 for FastWAM 50k. The paper reports reductions of 15.03% and 21.71%, respectively; PSNR rises from 31.441 to 31.663 dB. This is a matched prediction result across both branches.
Where the evidence stops. The evaluation is open-loop replay on box-task variants, without uncertainty estimates. Aggregate action error includes normalized padding channels; it should not be read directly as physical joint accuracy, task success, or broad scene generalization.
Table V(b). Observed gains are largest for coordination and tolerance to base displacement. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Translate the abbreviations before comparing rows: BD-SR is base docking, WBCM-SR is whole-body coordination, BDP-SR is base-displacement robustness, and AR-SR is autonomous recovery. Every value is a percentage from 79 trials per method, rounded to one decimal, and higher is better. Start with the FastWAM and DECOWAM rows because they provide the closest architectural comparison. Then inspect the other methods to see whether improvements concern merely reaching a dock or maintaining coordinated behavior under disturbances. These columns describe hardware capabilities; none is the final task-success rate reported separately in Table V(a). e13e14e15
What it supports. DECOWAM's coordination and displacement-robustness rates are 44.3% and 30.4%, versus FastWAM's 34.2% and 12.7%: gains of 10.1 and 17.7 percentage points. Recovery also rises from 27.8% to 32.9%. These observed margins are more pronounced than the final-task-success difference in the companion table.
Where the evidence stops. Coordination thresholds, displacement magnitudes, recovery windows, and uncertainty estimates are not specified. X-VLA, the strongest action-replay reference, is absent from this aggregate hardware table, so the table cannot establish hardware superiority over it.
Table V(a). Cumulative progress locates the stronger transport result while keeping final completion in perspective. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each row from approach through grasp, transport, and placement to task success. The caption states that every stage uses all attempts as its denominator, so the sequence shows accumulated attrition; the grasp percentage is not success conditional on approach. Each method has 79 attempts. Compare DECOWAM with FastWAM: the rates remain farther apart at transport than at the final stage, indicating that the intermediate advantage does not translate one-for-one into additional completed tasks. Read the left timing column separately from these cumulative rates. The accompanying text identifies DECOWAM's 58.2% as 46 completed tasks. e14e16
What it supports. DECOWAM reaches transport in 67.1% of attempts versus 59.5% for FastWAM, while final success is 58.2% versus 57.0%. Reported mean completion time decreases from 65 to 49 seconds. The result supports a stronger intermediate-stage profile, but only a small observed final-success advantage over the closest baseline.
Where the evidence stops. Mean completion time measures task duration; Table VI's evaluator latency is a separate quantity. The source gives no completion-time denominator, failure-time convention, or uncertainty, so 49 seconds must not be interpreted as controller inference speed.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Matched FastWAM replay prediction Fixed 23-episode Box-val slice, eight tasks; the same 16 replay batches, normalization, and inputs; eight 384 × 320 frames and 48 × 14 actions. | DECOWAM: 8.77e-4; 31.663 dB; 5.38e-5. Frame MSE ↓; PSNR (dB) ↑; normalized action MSE ↓ | FastWAM 50k: 1.032e-3; 31.441 dB; 6.87e-5. Authors report 15.03%/21.71% frame/action MSE reductions. Improvement is relative to the best observed 40k/50k/80k baseline checkpoint, under replay rather than executed-action evaluation. e10e11 |
| Structured-module replay ablation Separately trained suite; shared Stage-1 initialization and residual adapters; same 16-batch replay protocol. | Full: 8.09e-5. Normalized action MSE ↓ | Without quotient: 8.40e-5; without base velocity: 8.80e-5; adapter-only: 9.60e-5. The full suite model improves action MSE by 15.7% over adapter-only. Its reference differs from Table I; no GRL-only removal isolates the adversarial mechanism. e12 |
| Action-only reference comparison Table III, same 23-episode replay slice; VLA references predict actions without future RGB. | DECOWAM: 5.38e-5 / 4.01e-3 / 2.24e-2. Action MSE / MAE / mean Euclidean error ↓ | X-VLA: 2.11e-5 / 1.82e-3 / 1.00e-2. X-VLA has lower action error on all three measures. DECOWAM adds video prediction; this does not establish superior action prediction. e13 |
| Closed-loop task completion and transport Physical quadruped–arm platform; 79 trials per method; stage rates are cumulative over all attempts. | DECOWAM: 46/79 tasks (58.2%); 67.1% transport; 49 s. Task success / transport completion ↑; mean completion time ↓ | FastWAM: 57.0% success; 59.5% transport; 65 s. Final success is comparable to FastWAM; the transport and timing differences are larger. No uncertainty estimates or completion-time aggregation details are provided. e14 |
| Real-robot whole-body coordination and robustness Table V(b), 79 trials per method; reported percentages rounded to one decimal. | DECOWAM: 44.3% / 30.4% / 32.9%. Coordination / base-displacement robustness / autonomous recovery success ↑ | FastWAM: 34.2% / 12.7% / 27.8%. Observed coordination and displacement-robustness gains are 10.1 and 17.7 percentage points. Operational thresholds and perturbation magnitudes are not specified. e15 |
| Stage-2 adaptation and evaluator cost Table VI checkpoint diagnostics; evaluator hardware and sampling configuration not reported. | DECOWAM: 25.95M trainable; 6751.38M total; 1333.2 ms. Trainable parameters; total parameters; evaluator latency | FastWAM: 6020.75M trainable; 6725.44M total; 1196.6 ms. Approximately 232-fold fewer Stage-2 trainable parameters, with 11.4% higher evaluator latency. This excludes prior alignment cost and does not imply a smaller deployed model. e04e16 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table II, continued on p. 6. The added conditioning paths improve a separately trained ablation reference beyond residual adapters alone. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the first row as the reference for this table, not the main result in Table I: Section VI.B explicitly says this suite was trained separately. All variants retain residual adapters and share Stage-1 initialization. The quotient removal deletes the future-information bottleneck. The base-velocity removal deletes explicit ego-motion conditioning. The final w/o decoupled row removes all three structured paths, leaving adapter-only adaptation; it is not an isolated removal of the dual latent. Scan rightward to separate visual metrics from action metrics. Every removal worsens all five displayed diagnostics, but the magnitudes differ across variants and outputs. e05e06e07e12
What it supports. Within this suite, full-model action MSE is 8.09e-5 compared with 9.60e-5 for adapter-only, a reported 15.7% reduction. Removing the quotient or base-velocity condition yields 8.40e-5 or 8.80e-5. These rows support benefits from structured adaptation beyond the shared residual adapters on this replay protocol.
Where the evidence stops. No row removes only gradient reversal or only the base/arm factorization. The table therefore cannot isolate the adversarial separation mechanism, and its separately trained full row must not be substituted for the headline result.
7. Analysis & limitations
7.1 What the evidence leaves open
The dataset bookkeeping is unresolved: Section V describes 795 raw HDF5 episodes before reporting 1,487 filtered episodes, without explaining that transition. The 217/214-episode converted subsets and replay slice also require immutable membership manifests to audit separation. e09
The broader WAM table mixes eight-frame 384p outputs with four-frame 160p or 128p outputs. Aggregate normalized 14-D action errors include loader-padding channels and do not directly express physical control error. e03e10e13
Hardware comparisons and representative images do not isolate the video branch's causal contribution. The ablation removes several paths together in its adapter-only row; there is no standalone GRL ablation, repeated-seed uncertainty, or matched hardware module-ablation table. e12e14e15e17
7.2 Questions for discussion
- Would a capacity-matched, nonadversarial two-latent control retain the reported coordination benefit?
- Do gains persist when replay is stratified by base speed and evaluated in unnormalized arm and base units?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the exact ARMDOG conversion, checkpoint, split membership, normalization, language embeddings, evaluator, and robot interface. The source specifies timestamp interpolation and short-dropout repair but no exact repair thresholds; the data release is described as planned. e04e09e10
Missing controls include optimizer, learning rate, batch size, an unambiguous Stage-2 schedule, bottleneck subloss weights, ego-token scale, flow solver/step count, timing hardware, execution cadence, and hardware success thresholds. Proposed checks should compare matched seeds and report arm/gripper, base, and padding errors separately. e04e05e07e08e10e14e15e16
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 correctly paired privileged future information improve the deployable student?
Reader-proposed check, not an executed experiment: fix the 50k initialization, adapters, split membership, action normalization, loss weights, and compute budget. Across matched seeds, train the bottleneck with correctly paired future summaries versus summaries shuffled between training examples, retaining the original current inputs and action targets. Verify episode-level separation first, and evaluate only the causal student path on identical replay batches. Report arm/gripper and base errors separately, together with teacher–student distance. The privileged-information explanation predicts a repeatable advantage from correctly paired futures. Similar student performance would weaken the claim that accurate future supervision, rather than extra auxiliary optimization, explains the gain. e04e05e09e10e12
Check 2: Is gradient reversal responsible for factor separation and coordination?
Reader-proposed check, not an executed experiment: retain both 16-D factors, all heads, residual adapters, bottleneck, and ego-motion condition. Compare active gradient reversal with a control that stops cross-task gradients at the factor encoders while continuing to train the cross heads. Match initialization, batches, seeds, and optimization budget. Freeze each trained representation and fit held-out probes for assigned and opposite action subsets; also measure per-factor action error and static-base drift under a fixed, explicitly documented robot protocol. The proposed mechanism predicts reduced opposite-factor predictability without losing assigned-factor accuracy, accompanied by better coordination. Unchanged probe leakage or coordination would challenge a GRL-specific explanation. e06e07e12e15e17
8.3 Reading coverage
Visual audit: All eight original PDF pages were visually inspected: title/authors/version on p. 1; Figure 1 on p. 2; method equations on pp. 3–4; Figure 2 on p. 4; Figure 3 and Table I on p. 5; Table II caption on p. 5 and rows on p. 6; Tables III–IV, Figure 4, and Table V(b) on p. 6; Table V(a), Figure 5, and Table VI on p. 7; references on p. 8. All six final original crops were inspected for legibility and complete graphic/table content. The edition uses Figures 1–2 and Tables I, II, V(b), and V(a). No appendix appears in this PDF; separate supplements remain unverified.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. 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
- PDF p. 1: title, authors, version stamp, abstract, and Section I Introduction
- PDF pp. 2–3: Sections I–II and III Background: FastWAM World–Action Backbone
- PDF pp. 3–4: Section IV.A–D, Problem Formulation, Staged Parameter-Efficient Adaptation, Decoupled Conditional Interfaces, Training Objective and Deployment; Eqs. (1)–(19)
- PDF pp. 4–5: Section V, The ARMDOG Dataset
- PDF pp. 5–7: Section VI.A–D, replay protocol, main comparisons, ablations, real-robot experiments, and deployment diagnostics
- PDF pp. 7–8: Section VII Conclusion and complete References
Outside the original text pass
- Identity/version: the title page identifies arXiv:2608.20114v2 [cs.AI], 21 August 2026. Its title and seven authors match the supplied catalog. Earlier revisions were not supplied, so revision differences cannot be assessed.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The extraction's visual omission was addressed by inspecting all eight original PDF pages, every figure and table, and all six final crops.
- Separate supplemental material availability has not been fully verified.
- No appendix is present in the supplied eight-page PDF. External references, code, data files, and separate supplements were not inspected; experiments were not reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block and left-margin arXiv stamp
The supplied PDF is DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation, arXiv:2608.20114v2, 21 August 2026. Authors: Siyuan Ma, Boshi Zhang, Yutian Zhang, Qinglian Wu, Jiaqi Zhai, Dong Wei, Qiaojun Yu. Four affiliations and contribution/correspondence marks are shown.
Go to primary source ↓e02PDF p. 1, Section I(a)–(c)
Moving viewpoints mix ego-motion and scene motion; arm control is described around 15–30 Hz and base commands around 3–5 Hz. Base velocity is both a predicted control target and a visual condition.
Go to primary source ↓e03PDF p. 2, Figure 1; p. 3, Sections III and IV.A, Eqs. (1)–(2)
FastWAM pairs a WAN video expert with ActionDiT. The conditional joint model predicts eight frames and 48 actions with arm/gripper channels [0:7], base [7:10], and padding [10:14].
Go to primary source ↓e04PDF p. 3, Section IV.B, Eqs. (3)–(5); p. 5, Section VI.A(a)
Stage 1 aligns FastWAM for 50k steps; Stage 2 freezes the aligned checkpoint and trains four added groups. Adapters have a 128-D bottleneck. Trainable counts are 6020.75M versus 25.95M.
Go to primary source ↓e05PDF pp. 3–4, Section IV.C(a), Eqs. (6)–(11), Figure 2; p. 5, Section VI.A(a)
Teacher input includes current/future latent summaries and state; student omits future data. Only the student conditions actions. Action reconstruction, stopped-gradient distillation, and pairwise geometry matching train the 64-D bottleneck; geometry scales use batch medians.
Go to primary source ↓e06PDF p. 4, Section IV.C(b), Eqs. (12)–(14)
Separate 16-D base and arm factors bias action context. Direct prediction heads preserve their assigned channels; gradient reversal reverses cross-task gradients entering the factor encoders.
Go to primary source ↓e07PDF p. 4, Section IV.C(c), Eqs. (15)–(16)
Current normalized base velocity (v_x, v_y, ω_z) is projected and added to every WAN video token. The source explicitly distinguishes this condition from geometric warping.
Go to primary source ↓e08PDF p. 4, Section IV.D, Eqs. (17)–(19) and deployment paragraph
Action and video experts use conditional flow matching. Loss weights are λ_v=λ_a=1, λ_q=0.2, λ_ba=0.1; γ_q, γ_ba, η_q, η_ba are one. Teacher and auxiliary heads are removed at inference; both flows use current image, state, and language.
Go to primary source ↓e09PDF p. 2, Section I(d); pp. 4–5, Section V and Figure 3
Converted snapshot: 217 episodes, 27 folders, 56,041 frames. Section V cites 795 raw episodes, then 1,487 filtered episodes from five folders, 343,550 frames, 321.3 minutes at 15 Hz. Stage 2: 214 episodes/26 tasks excluding Legacy val; Box-val: 23 episodes/eight tasks/4,323 frames. Four task-family shares are 56%, 39%, 4%, 1%. Release is planned; raw-to-filtered count reconciliation and split membership are not given.
Go to primary source ↓e10PDF p. 5, Section VI.A, Replay Protocol, Models and Metrics
Box-val contains box_move, box_soft, and box_stay variants. FastWAM-family inputs, normalization, and evaluator are shared over 16 batches. Outputs are eight 384 × 320 frames and 48 × 14 actions; action metrics operate on normalized 14-D vectors.
Go to primary source ↓e11PDF p. 5, Table I, all rows and following paragraph
FastWAM 40k/50k/80k action MSE is 1.154e-4/6.87e-5/4.77e-4. DECOWAM has frame MSE 8.77e-4, PSNR 31.663, action MSE 5.38e-5 versus 50k values 1.032e-3, 31.441, 6.87e-5. Reported reductions are 15.03% and 21.71%.
Go to primary source ↓e12PDF p. 5, Section VI.B(a) and Table II caption; p. 6, Table II rows and following paragraph
Separately trained full ablation model: frame MSE 9.35e-4, PSNR 31.378, SSIM 0.99241, action MSE 8.09e-5, MAE 4.310e-3. Removing quotient/base-velocity/all structured paths gives action MSE 8.40e-5/8.80e-5/9.60e-5. All rows retain residual adapters; no individual GRL-removal row is present.
Go to primary source ↓e13PDF p. 6, Tables III–IV, RGB and action-error columns
X-VLA leads Table III with MSE 2.11e-5, MAE 1.82e-3, L2 1.00e-2 versus DECOWAM 5.38e-5, 4.01e-3, 2.24e-2. Table IV reports DECOWAM leading all listed metrics, but X-WAM/UVA use four-frame 160p/128p outputs versus its eight-frame 384p interface.
Go to primary source ↓e14PDF p. 6, Section VI.C; p. 7, Table V(a), caption and DECOWAM/FastWAM rows
Each method receives 79 physical trials under common observations, low-level control, and safety constraints. Cumulative DECOWAM approach/grasp/transport/placement/success is 92.4/69.6/67.1/58.2/58.2%, with 46 completions and 49 s mean time. FastWAM has 91.1/63.3/59.5/57.0/57.0% and 65 s. Uncertainty and completion-time aggregation rules are absent.
Go to primary source ↓e15PDF p. 6, Table V(b), caption, DECOWAM/FastWAM rows and abbreviation definitions
BD-SR, WBCM-SR, BDP-SR, AR-SR mean docking, coordination, displacement robustness, and autonomous recovery success. DECOWAM: 87.3/44.3/30.4/32.9%; FastWAM: 83.5/34.2/12.7/27.8%. Rates use 79 trials; detailed thresholds and perturbation magnitudes are not specified.
Go to primary source ↓e16PDF p. 7, Section VI.D and Table VI
DECOWAM/FastWAM total parameters: 6751.38M/6725.44M; trainable parameters: 25.95M/6020.75M; evaluator latency: 1333.2/1196.6 ms. Text reports about 232-fold fewer Stage-2 updates and 11.4% more latency; timing hardware and solver settings are not given.
Go to primary source ↓e17PDF p. 6, Figure 4 and caption; p. 7, Figure 5, caption and Section VI.C discussion
Figure 4 is one replay sample with model-native visual interfaces. Figure 5 shows representative X-VLA/FastWAM failures and DECOWAM success during a static-base reach. The authors attribute baseline motion to motion-heavy training; these examples do not constitute an isolated causal test.
Go to primary source ↓8.5 Primary sources
DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation ↗
PDF · 5,922 extracted words
Source fingerprint
4eadfbcfa5af90428082b4adba0bb71e045bca9d0b518b2947a4a00a4cdc0f23