Do World Action Models Generalize Better than VLAs? A Robustness Study
1. Paper overview
In one sentence: Video-based policies show strong visual robustness, but benchmark dependence, unresolved score aggregation and costly inference prevent a universal WAM advantage. e05e06e10e11e20
| At a glance | What to know |
|---|---|
| Research problem | Source description Does video-based future prediction make robot policies robust when familiar manipulation tasks undergo visual, language or initial-state changes? The study evaluates executed simulated actions; future-image examples supply separate diagnostics. Its contribution is a benchmark extension and comparative study. e01e05e13 |
| Core mechanism | |
| A key reported result | RoboTwin 2.0-Plus: visual versus geometric robustness: LingBot-VA: light 89.0%, noise 80.9%, layout 87.9%, camera 28.9%, robot initial state 36.2%. Branch success rate (%). Individual perturbation branches in Table 3. π0.5 camera: 45.6%; MOTUS robot initial state: 85.0%. Strong visual robustness coexists with geometric sensitivity; no policy family wins every branch. e10 |
| Reading caution | Reader analysis Total aggregation is unresolved. Reader arithmetic gives LingBot-VA 71.64% across seven perturbation columns, or 74.20% including Original; Fast-WAM gives 72.67% across seven, or 74.99% including Original. The reported 74.2/72.7 pair mixes apparent conventions. Raw episode records are needed before treating this pair as a consistent ranking. e06e10e20 |
Core contributions
Figure 1. RoboTwin 2.0-Plus changes the conditions surrounding a familiar task. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the upper-left clean scene, then compare each labeled panel with it. Camera changes the head-camera view; Robot changes the initial joints and grippers; Language changes the instruction while preserving the task intent. The lower row changes illumination, background appearance, image corruption and object layout. The codes point to Appendix A, rather than denoting severity levels. Camera uses C1+C3 because C2 is disabled by default. Layout includes both O1 distractors and O2 target-pose changes, so its score cannot be interpreted as clutter resistance alone. These panels illustrate separate evaluation branches, not one episode containing all seven dimensions. e05e06e07
What it supports. The experimental unit is a task under a specified perturbation branch. Appendix A prescribes a clean baseline and seven branches, each with 50 episodes per task across 50 tasks. This design supports diagnosis by perturbation type and makes the branch configuration as important as the model name.
Where the evidence stops. The figure shows examples, not the distribution of severity or success. Sub-dimensions can be combined inside a branch, and the disabled spherical-camera perturbation limits what the default camera result covers.
2. Motivation
2.1 The problem and the proposed response
Does video-based future prediction make robot policies robust when familiar manipulation tasks undergo visual, language or initial-state changes? The study evaluates executed simulated actions; future-image examples supply separate diagnostics. Its contribution is a benchmark extension and comparative study. e01e05e13
2.2 What this reading follows
This study asks how robot policies behave when familiar tasks are presented through altered images, instructions or initial configurations. It introduces RoboTwin 2.0-Plus for bimanual manipulation and also evaluates the single-arm LIBERO-Plus benchmark. The most useful evidence is the pattern across perturbation types: video-based policies can resist noise and lighting changes yet remain sensitive to camera geometry and robot state. The headline totals need care. π0.5 leads the reported LIBERO comparison, and RoboTwin’s Total column has an aggregation inconsistency. Read the figures as a guide to what was tested, which mechanisms are plausible, and what a controlled reproduction must resolve. e05e06e10e11e20
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 | Benchmarks & simulators |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The benchmark/simulator and robustness-evaluation classification is supported. Architecture, prediction paradigm and quadrant remain not applicable to the study as a whole. Evaluated policies span unified and multi-transformer designs, joint prediction and inverse dynamics; these properties cannot be assigned to the benchmark itself. e02e03e05e06
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 Separate future prediction from the route that produces actions
Section 2.3 supplies the conceptual distinction behind the comparison. In its notation, a direct VLA maps current state h_t to action a_t. A joint WAM predicts both the next visual state h_{t+1} and action from h_t. An inverse-dynamics-style design instead predicts a future first and conditions action generation on current and future states. That ordering matters: training on future images does not automatically make those predictions necessary during control. Fast-WAM can generate actions without future images at inference, whereas LingBot-VA conditions actions on predicted visual state. Cosmos-Policy uses latent-frame diffusion for multiple prediction targets. Reader interpretation: the meaningful comparison has several axes—backbone prior, embodied training, modality interaction and inference dependency. Treating all of them as a single WAM switch would hide the mechanism being tested. e02e03e04e14
Table 1. The WAM label spans different architectures and prediction orderings. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Backbone first, then the four design columns. MOT means separate video and action transformers interacting through attention; it does not merely mean multiple outputs. Pretrain Free means no task-agnostic embodied pretraining, while video-backbone pretraining and task-specific learning still remain. Causal Pred. covers either future-to-action or action-to-future conditioning; it does not tell the direction by itself. Section 2.2.3 identifies LingBot-VA as future-conditioned action generation and GigaWorld-Policy as action-conditioned future generation. AR Gen. marks autoregression. Keep the question mark in LingBot-VA’s MOT cell: the caption explicitly says the released model uses a unified transformer instead of the originally described mixture. e02e03e04e09e19
What it supports. Architecture, learning stage and prediction ordering are independent comparison axes. Tables 1–2 identify both Cosmos-Policy and Fast-WAM as omitting embodied pretraining. The paper’s stronger RQ1 statement that Fast-WAM is the only evaluated model with this property is inconsistent with those tables.
Where the evidence stops. This is the study’s summary of other methods, not independent inspection of their code. DreamZero and GigaWorld-Policy appear in the taxonomy but are excluded from quantitative evaluation in this revision; a taxonomy row does not imply a measured result.
5.2 Reconstruct the evaluation denominator before interpreting superiority
Appendix A makes the intended RoboTwin experiment concrete: each of fifty tasks receives a clean configuration and seven perturbation branches, with fifty episodes per configuration. Within a branch, some interventions are combined; layout includes target-pose changes as well as distractors. This determines what each column can diagnose. It also makes Total aggregation checkable. From the printed Table 3 cells, LingBot-VA averages 71.64% over seven perturbation columns, while its reported 74.2% matches including Original. Fast-WAM’s reported 72.7% instead matches the seven-column mean. Reader analysis: those totals cannot be treated as an evidently common robustness metric. Recover per-episode outcomes, separate clean from perturbed trials and use one stated aggregation rule before assessing the close comparison between those two policies. e06e07e10e20
5.3 Distinguish a promising explanation from a controlled result
Fast-WAM offers an informative contrast: its RoboTwin checkpoint uses clean and randomized demonstrations, whereas the LIBERO checkpoint uses clean demonstrations alone. The paper attributes the different robustness profiles to training-data diversity and suggests a role for prediction ordering. Those hypotheses are plausible, but the benchmark, robot and observation interface also change. Figure 3 provides another candidate mechanism—background distortions in generated futures—without quantifying its effect on action success. Reader analysis: neither observation isolates a causal explanation. A deployment argument needs a further separation between success and timing. Table 5 varies denoising schedules and action chunk sizes, and Fast-WAM’s timing comes from different hardware. Jointly reporting controlled success and response-time measurements would show whether a robust policy remains useful at the intended execution cadence. e05e09e12e13e14e15
5.4 Training and inference
During training
Backbone pretraining precedes task-agnostic embodied pretraining and task-specific finetuning; some policies omit the embodied stage. This still uses pretrained representations and robot demonstrations. Tables 1–2 identify both Cosmos-Policy and Fast-WAM as omitting that stage, despite RQ1 describing Fast-WAM as uniquely doing so. e02e04e19
JAX π0.5 is finetuned on 27.5k RoboTwin demonstrations for 60k steps, batch 64, with AdamW (β₁=0.9, β₂=0.95), clipping 1.0 and cosine learning-rate decay from 2.5×10⁻⁵ to 2.5×10⁻⁶ using delta joint actions. Fast-WAM uses randomized-plus-clean RoboTwin versus clean-only LIBERO data. e08e09
During inference
Cosmos-Policy represents policy/world/value quantities within latent-frame diffusion; LingBot-VA conditions action on predicted visual state. Fast-WAM can omit visual generation during inference although state and action are jointly trained. Auxiliary future prediction is thus distinct from mandatory test-time imagination. e02e03e14
Rollout success and chunk-generation time are separate measurements. LingBot-VA timing settings use 3/5 state/action denoising steps for RW and 25/50 for RT. The RW label denotes a runtime configuration, not new physical-robot robustness evidence in this study. e05e14
5.5 Implementation flow
- Separate policy families
Direct VLAs predict actions from current state. WAMs jointly predict future state and action, condition actions on a predicted future, or predict action before future state. MOTUS is a hybrid because action generation uses a separate VLM expert. These are this source’s descriptions of existing models. e02e03
- Perturb one dimension per branch
Run one clean configuration and seven branches, each with 50 episodes per task. Sub-dimensions coexist within branches: layout changes distractor count and target pose. C2 spherical viewpoint changes are disabled, leaving 20 of 21 implemented sub-dimensions active by default. e06e07
- Keep interfaces distinct
LIBERO uses a single Franka arm, two cameras and delta end-effector control; RoboTwin uses dual Aloha-AgileX arms, three cameras and joint control. Cross-benchmark differences therefore change embodiment, sensing and simulator. RoboTwin comparisons use one model across tasks. e05e10
6. Experiments & results
This study extends RoboTwin 2.0 with controlled perturbations and compares video-based world action models with VLAs on two simulated manipulation benchmarks. WAMs have strong visual-perturbation results, but π0.5 leads the reported LIBERO-Plus total. Unequal training histories, slow inference and an unresolved RoboTwin aggregation inconsistency limit broad claims about architectural superiority.
This source is a comparative benchmark study and supplies no new network architecture diagram or controlled removal-of-component ablation. Table 1 provides its architecture comparison and Figure 1 its experimental method illustration. Figure 3 serves as the diagnostic visual in the ablation section, explicitly labeled qualitative. Appendix A enables ablation-style configurations, but RQ2’s checkpoint comparison changes benchmark and training data together; it is not a matched architectural or data ablation. e02e05e06e12e13
6.1 Read the original evidence
Table 3. LingBot-VA is strong on visual perturbations; the published Total column needs reconciliation. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First separate Original, the RoboTwin Easy baseline, from the seven perturbation columns. Read across a row to see a policy’s vulnerability, then down a column to compare policies under that intervention. Bold marks the best reported entry and underlining marks second place. LingBot-VA’s strong light, background, noise and layout entries contrast with its low camera and robot-state entries. The Total column should be read as a published statistic with an unresolved definition. Its 74.2 for LingBot-VA matches an average including Original, while Fast-WAM’s 72.7 matches an average excluding Original. Appendix A specifies equal episode counts across configurations. e06e10e20
What it supports. LingBot-VA achieves 89.0% under light changes, 80.9% under noise and 87.9% under layout changes. Yet its camera score is 28.9%, below π0.5’s 45.6%, and its robot-state score is 36.2%, below MOTUS’s 85.0%. These branch results support a selective robustness advantage.
Where the evidence stops. The seven perturbation columns average 71.64% for LingBot-VA and 72.67% for Fast-WAM by reader calculation. This differs from the reported Total ranking. Keep the published cells intact and resolve aggregation from episode records before asserting that ranking.
Table 4. A second benchmark changes the overall ordering and exposes checkpoint dependence. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read this as a separate single-arm evaluation rather than combining its percentages with RoboTwin. Original is clean LIBERO, while Camera through Layout are perturbed conditions. The row groups distinguish VLAs, a VLA with world-model learning, and WAMs. π0.5’s reported Total is 85.7, above Cosmos-Policy’s 82.2 and GE-Act’s 80.3. Fast-WAM has high clean success but weak camera, background and noise performance. The caption matters: several VLA rows come from prior papers, while the study reruns other released checkpoints; π0 rerun and π0.5 use JAX. Thus the table combines evidence with different evaluation provenance. e05e09e11e12e20
What it supports. The reported LIBERO ordering rules out a blanket claim that WAMs outperform VLAs. Fast-WAM falls from 97.6% Original to 51.5% reported Total with its clean-only checkpoint. The corresponding RoboTwin checkpoint uses randomized training, making data diversity a plausible explanation worth testing under controlled conditions.
Where the evidence stops. Benchmark, robot and sensing change alongside Fast-WAM’s training data, so this is not an isolated data-diversity ablation. Total weighting and uncertainty are unspecified. Preserve the source’s reported totals without presenting them as independently recomputed averages.
Table 5. Robustness comes with configuration-dependent inference cost. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the action chunk size before comparing wall-clock time: each column predicts a different amount of action output. π0.5 takes 63 ms, GE-Act 300 ms and Cosmos-Policy 390 ms in the study’s timing comparison. LingBot-VA has two configurations: RW uses three state and five action denoising steps, while RT uses twenty-five and fifty. Their corresponding times are 480 and 5230 ms. The bottom row divides chunk latency by π0.5 latency; it is not an equal-horizon control-throughput measure. Retain the starred Fast-WAM footnote: its 190 ms was imported from different hardware despite the caption’s general same-device statement. e14e15e18
What it supports. State-generation choices strongly affect the reported runtime tradeoff. The measured GE-Act and Cosmos-Policy settings are 4.8 and 6.2 times π0.5 per chunk, while LingBot-VA RT is 83.0 times. Fast-WAM’s nominal 3.0-times figure has different measurement provenance and should not be treated as a hardware-controlled speedup.
Where the evidence stops. The device is unnamed, and Table 9 lists different chunk sizes for several methods without reconciling configurations. The conclusion’s blanket 4.8-times minimum also omits Fast-WAM’s starred exception. These timings do not measure end-to-end physical control performance.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0-Plus: reported overall success 50 simulated bimanual tasks; original Easy baseline and seven perturbation branches; one model across tasks. | LingBot-VA 74.2%; Fast-WAM 72.7%. Reported Total success rate (%) | MOTUS 71.5%; π0.5 58.6%; X-VLA 53.1%. These published totals do not follow a common seven-branch mean; the apparent LingBot-VA lead over Fast-WAM requires aggregation reconciliation. e06e10e20 |
| RoboTwin 2.0-Plus: visual versus geometric robustness Individual perturbation branches in Table 3. | LingBot-VA: light 89.0%, noise 80.9%, layout 87.9%, camera 28.9%, robot initial state 36.2%. Branch success rate (%) | π0.5 camera: 45.6%; MOTUS robot initial state: 85.0%. Strong visual robustness coexists with geometric sensitivity; no policy family wins every branch. e10 |
| LIBERO-Plus: reported overall success Four single-arm task suites; Table 4 mixes study evaluations and attributed literature results. | Cosmos-Policy 82.2%; GE-Act 80.3%. Reported Total success rate (%) | π0.5 85.7%; VLA-JEPA 77.9%. π0.5 leads this table. Original success is 98.5% for Cosmos-Policy and 96.9% for π0.5. Total weighting remains unexplained. e05e11e20 |
| Fast-WAM: checkpoint-dependent robustness Randomized-plus-clean RoboTwin training versus clean-only LIBERO training. | RoboTwin 91.2 / 72.7; LIBERO 97.6 / 51.5. Original / reported Total success (%) | LIBERO suite totals: Object 71.2%, Goal 39.2%. High clean success does not guarantee robustness. Data diversity is plausible, but changing benchmark and robot prevents a controlled causal conclusion. e09e12e16 |
| Per-chunk inference cost Table 5 study-device timings; Fast-WAM is a separate-hardware literature timing. | GE-Act 300; Cosmos-Policy 390; LingBot-VA RW 480 and RT 5230. Wall-clock milliseconds per action chunk | π0.5 63 ms; Fast-WAM 190 ms on different hardware. Measured WAM configurations cost at least 4.8 times π0.5 per chunk. Fast-WAM’s nominal 3.0-times ratio is not a same-device comparison; unequal chunk sizes prevent equivalent-throughput conclusions. e14e15 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 3. Future-image predictions reveal useful denoising and a background-sensitive failure mode. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each perturbation group vertically: GT is the ground-truth future image and Pred. is Cosmos-policy’s generated future. Each group includes a third-person view and a wrist view, so inspect both before judging the prediction. Under Noise, compare the blurred moving arm with its sharper predicted counterpart. Under Light, inspect arm and object placement rather than assuming that similar brightness guarantees correct dynamics. Under Background, compare the patterned surface, robot geometry and object contours; the prediction visibly changes scene structure and appearance. These are paired future-state illustrations, not a sequence of executed actions or a numerical image-quality evaluation. e13
What it supports. The examples are consistent with the authors’ hypothesis that a video backbone can preserve useful motion structure under some corruptions while distorting unfamiliar backgrounds. They provide a concrete candidate failure mechanism to investigate alongside the success-rate tables, rather than a measured explanation of those tables.
Where the evidence stops. This is a qualitative diagnostic, not a controlled ablation. No quantitative future-error metric or intervention connects the displayed distortions to subsequent action failure. Attractive or denoised predictions alone do not establish reliable control.
7. Analysis & limitations
7.1 What the evidence leaves open
Total aggregation is unresolved. Reader arithmetic gives LingBot-VA 71.64% across seven perturbation columns, or 74.20% including Original; Fast-WAM gives 72.67% across seven, or 74.99% including Original. The reported 74.2/72.7 pair mixes apparent conventions. Raw episode records are needed before treating this pair as a consistent ranking. e06e10e20
No matched backbone/data ablation establishes that video priors cause robustness or are necessary. Training histories and imported LIBERO rows differ. Main success tables lack confidence intervals and seed variability. Simulated task perturbations do not establish physical deployment or unrestricted novel-task generalization. e04e05e09e10e11e12
Figure 3 gives selected future-prediction successes and distortions, without quantitative prediction-error or mediation tests linking images to action failure. Figure 2 likewise illustrates selected rollout outcomes. e13e17
7.2 Questions for discussion
- Would matched training-data ablations preserve the apparent WAM advantage?
- How should Total aggregate clean and perturbed trials consistently?
- Which denoising and execution horizons preserve robustness under a control-time budget?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs checkpoint variants, benchmark configurations and per-episode success records. Preserve C2 exclusion, deterministic noise cycling and combined layout perturbations. Simulators are named but exact builds, a complete seed list, task-timeout/success-checker details and timing-device identity are not specified. e05e06e07e08e09e14
Resolve configuration differences before timing: Table 5 versus Table 9 gives chunk sizes 36/54 for GE-Act, 16/48 for MOTUS and 32/4 for LingBot-VA. Proposed checks should hold data and hardware fixed and distinguish predicted horizons from executed actions. e14e15
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: Recompute clean and perturbed totals from episode outcomes
Reader-proposed check, not an executed experiment: obtain the evaluated LingBot-VA and Fast-WAM checkpoint identifiers and per-episode RoboTwin outcomes. Verify fifty tasks and fifty episodes for each of eight configurations, including failed or aborted episodes. Compute each branch success rate, a seven-branch perturbation mean and a separately labeled eight-configuration mean; use the same rule for both models. Recompute the table from those records and report uncertainty across tasks. The falsifiable target is whether one documented weighting reproduces both published Total entries. If only different inclusion rules reproduce 74.2 and 72.7, the headline ordering is not a consistent aggregate comparison. e06e09e10e20
Check 2: Isolate training diversity within one Fast-WAM benchmark
Reader-proposed check, not an executed experiment: on RoboTwin, finetune the same Fast-WAM initialization with clean-only versus clean-plus-randomized demonstrations, matching task coverage, demonstration count and optimization steps. Keep evaluation seeds, default C2 exclusion, action interface and inference settings fixed. Evaluate all seven perturbation branches and the clean baseline using the reconciled aggregation rule; run a matched π0.5 data comparison as a control for a general augmentation effect. The hypothesis predicts reduced perturbation degradation with diverse data beyond a mere clean-success change. If matched conditions erase the benefit, the cross-benchmark explanation weakens. Measure latency on identical hardware with the executed action horizon documented to keep accuracy gains separate from timing changes. e07e08e09e12e14e15
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page; Table 1 and its architecture caveat; method and prediction formulations; training data and finetuning details; Figures 1–3; main result and runtime tables; conclusion; and all appendix protocol, parameter, benchmark, model-configuration and suite-result pages. All six final original crops were separately viewed, with labels and the timing footnote retained. The model-design cells were cross-checked against their caption and Sections 2.2.3–2.3; no architecture graph was supplied. Pages 2–3 and references 13–17 were read in the complete text pass but not visually inspected. External code and separate supplements remain outside this review.
PDF pages inspected for this edition: 1, 4, 5, 6, 7, 8, 9, 10, 11, 12, 18, 19, 20, 21. 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 and title metadata (p. 1)
- 1 Introduction (pp. 1–2)
- 2 Related Works, including 2.1–2.3 (pp. 2–6)
- 3 Experiments, including 3.1–3.3 and RQ1–RQ4 (pp. 7–11)
- 4 Conclusion (p. 12)
- References (pp. 13–17)
- Appendix A, A.1–A.3 (pp. 18–19)
- Appendix B (pp. 19–20)
- Appendix C and C.1 (pp. 20–21)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This acquisition limitation was addressed by visually inspecting the relevant original pages and final crops.
- Separate supplemental material availability has not been fully verified.
- Reviewed version: arXiv:2603.22078v5, 30 July 2026. The catalog submission date is 23 March 2026; earlier editions were not supplied, so revision changes cannot be reconstructed. Observed title and all 14 authors match the catalog.
- All eight text chunks were read. Visual inspection covered pages 1, 4–12 and 18–21; pages 2–3 and reference pages 13–17 were read as text only.
- Code, external model papers and linked resources were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block and arXiv margin
Title and all 14 authors match the catalog. The artifact is arXiv:2603.22078v5 [cs.RO], 30 July 2026. Affiliations are Huawei Technologies and University of Toronto.
Go to primary source ↓e02PDF pp. 4–5, Table 1 and Section 2.2.3
WAMs adapt pretrained video backbones for action generation. Table 1 distinguishes mixture-of-transformers, embodied pretraining, causal prediction and autoregression. Its caption says released LingBot-VA uses a unified transformer despite the architecture in its original paper; the MOT cell retains a question-marked check. MOTUS uses a separate VLM action expert.
Go to primary source ↓e03PDF p. 6, Section 2.3, prediction-scheme paragraph
The paper formulates direct action prediction, joint future-state/action prediction, and a future predictor followed by action generation conditioned on current and predicted future states. It also describes action-conditioned future prediction.
Go to primary source ↓e04PDF p. 6, Table 2 and Section 2.3
Training distinguishes video/VLM backbone pretraining, task-agnostic embodied pretraining, optional embodied post-training and task-specific finetuning. Table 2 lists diverse robotic/web objectives for pi0.5 and no embodied-pretraining data for Cosmos-Policy or Fast-WAM.
Go to primary source ↓e05PDF p. 7, Figure 1 and Section 3.1; p. 20, Table 8
RoboTwin 2.0-Plus extends 50 bimanual Aloha-AgileX tasks with seven perturbation dimensions, three cameras and joint control in SAPIEN. LIBERO has 40 single-arm Franka tasks, two cameras and delta end-effector control in MuJoCo.
Go to primary source ↓e06PDF p. 18, Appendix A and Sections A.1–A.2; p. 20, Table 7
The protocol uses 50 tasks, eight configurations per task and 50 episodes per configuration: clean plus seven branches, each changing one dimension. A branch can combine sub-dimensions.
Go to primary source ↓e07PDF pp. 18–19, Section A.3 and Table 6
Twenty-one sub-dimensions are implemented, twenty active by default. C2 spherical camera variation is disabled to avoid instability. Noise types cycle by episode. Layout combines 3–15 distractors with 2 cm standard-deviation target-position noise and yaw up to 15 degrees. There are 2,500 instruction variants.
Go to primary source ↓e08PDF p. 7, Section 3.2, pi0.5 finetuning paragraph
JAX pi0.5 is finetuned on all 27.5k RoboTwin demonstrations for 60k steps, batch 64, AdamW beta1=0.9 and beta2=0.95, clipping 1.0, cosine learning-rate decay from 2.5e-5 to 2.5e-6, with delta joint actions.
Go to primary source ↓e09PDF p. 8, Section 3.2 continuation
Fast-WAM uses robotwin_uncond_3cam_384 with clean plus randomized RoboTwin demonstrations and a clean-only LIBERO checkpoint. DreamZero is excluded for checkpoint incompatibility and retraining/evaluation expense; GigaWorld-Policy lacks an available checkpoint in this revision.
Go to primary source ↓e10PDF p. 8, Table 3, all rows and caption
Reported RoboTwin Total values: LingBot-VA 74.2, Fast-WAM 72.7, MOTUS 71.5, pi0.5 58.6 and X-VLA 53.1 percent. Original is the Easy setting; one model is used across tasks. LingBot-VA light/noise/layout/camera/robot scores are 89.0/80.9/87.9/28.9/36.2; pi0.5 camera is 45.6 and MOTUS robot is 85.0.
Go to primary source ↓e11PDF p. 10, Table 4, all rows and caption
LIBERO reported Total: pi0.5 85.7, Cosmos-Policy 82.2, GE-Act 80.3, VLA-JEPA 77.9 and Fast-WAM 51.5 percent. Original: pi0.5 96.9, Cosmos-Policy 98.5 and Fast-WAM 97.6. Several rows are imported from other papers; pi0 rerun and pi0.5 use JAX.
Go to primary source ↓e12PDF p. 10, RQ2 Fast-WAM discussion; p. 8, Section 3.2
Authors attribute Fast-WAM differences to task-specific data diversity, calling the comparison a natural experiment. Original/Total scores are 91.2/72.7 for RoboTwin and 97.6/51.5 for LIBERO. Benchmark and embodiment also differ.
Go to primary source ↓e13PDF p. 11, Figure 3 and caption; pp. 9–10, RQ2 discussion
Ground-truth and Cosmos-policy future predictions are paired under noise, light and background changes in third-person and wrist views. Authors discuss arm denoising and background-induced spatial/color errors; the effect on action accuracy is offered as a likelihood.
Go to primary source ↓e14PDF p. 11, Table 5, footnote and RQ4
Per-chunk timings: pi0.5 63 ms, GE-Act 300 ms, Cosmos-Policy 390 ms, LingBot-VA RW 480 ms and RT 5230 ms. LingBot state/action denoising counts are 3/5 and 25/50. Fast-WAM 190 ms comes from different hardware despite the general same-device caption. The timing device is unnamed.
Go to primary source ↓e15PDF p. 11, Table 5; pp. 20–21, Table 9 and Appendix C
Table 5 chunk sizes for X-VLA/GE-Act/MOTUS/LingBot-VA are 30/36/16/32; Table 9 lists 32/54/48/4, repeated in Appendix C. Configurations are not reconciled. Table 9 marks inherited or unspecified entries with dashes.
Go to primary source ↓e16PDF p. 21, Table 10, Fast-WAM suite rows and caption
Fast-WAM reported LIBERO suite totals: Spatial 54.4, Object 71.2, Goal 39.2 and Long 41.1 percent. The caption repeats clean-only LIBERO versus randomized RoboTwin training.
Go to primary source ↓e17PDF p. 9, Figure 2 and caption
Three selected sequences show pi0.5 failures under zoom blur, layout and lighting perturbations, with LingBot-VA successes; these are qualitative examples.
Go to primary source ↓e18PDF p. 12, Section 4
The conclusion emphasizes visual robustness, camera/initial-state difficulties and inference cost. It repeats an at-least-4.8-times latency claim while also describing Fast-WAM as approximately three times slower.
Go to primary source ↓e19PDF p. 9, RQ1 Fast-WAM paragraph; p. 4, Table 1; p. 6, Table 2
RQ1 calls Fast-WAM the only evaluated WAM without embodied pretraining, but Tables 1–2 also identify Cosmos-Policy as omitting that stage.
Go to primary source ↓e20PDF p. 8, Table 3; p. 18, Section A.2; p. 10, Table 4
The protocol gives equal episode counts but Total aggregation is undefined. Reader arithmetic: LingBot-VA seven perturbation columns average 71.64%, or 74.20% with Original; Fast-WAM averages 72.67%, or 74.99% with Original. Table 4 likewise leaves Total weighting unexplained.
Go to primary source ↓8.5 Primary sources
Do World Action Models Generalize Better than VLAs? A Robustness Study ↗
PDF · 12,444 extracted words
Source fingerprint
f7606b9072527da9f0af9a87aeb7b950426b1004193183baab3840ecfd2b80f8