OA-WAM: Object-Addressable World Action Model for Robust Robot Manipulation
1. Paper overview
In one sentence: Persistent object addresses help the shared world/action model withstand geometric scene shifts, but the benefit depends on perception and does not establish universally robust control. e-probleme-architecturee-routinge-oa-ablatione-world-ablatione-reporting-gaps
| At a glance | What to know |
|---|---|
| Research problem | Author claim A language-conditioned robot must keep identifying the same target when viewpoint or scene arrangement changes. The authors argue that holistic world latents mix identity with context, making target selection unstable even when objects remain visible. OA-WAM tests whether an explicit, persistent addressing interface helps under these shifts. e-problem |
| Core mechanism | Source description Each object receives a frozen episode address and dynamic content. Parameter-free key masking and per-layer residual resets constrain the address channel within one shared multimodal trunk. e-architecturee-routing |
| A key reported result | LIBERO-Plus geometric robustness: Camera 80.5; robot 89.6; layout 82.8; Geo Avg 84.3. Success (%) and arithmetic mean across Camera/Robot/Layout. Zero-shot perturbations after standard-LIBERO finetuning; official generator, three seeds. Camera exceeds Cosmos-Policy 75.8 by 4.7 points; Geo Avg exceeds pi0.5 79.5 by 4.8. Robot trails X-VLA 89.7; layout trails pi0.5 85.7. The table supports best camera and geometric average among listed rows, not leadership on every geometric axis. e-robustnesse-evaluation |
| Reading caution | Source description Validation is simulator-only and uses privileged object poses. Small, reflective, transparent, occluded or blurred objects can defeat extraction; distractor invariance assumes weak physical coupling. The failure audit identifies dynamics, perception, engineering and ambiguous/blocked tasks rather than one universal cause. e-limitationse-preprocessing |
Core contributions
- Source description
Each object receives a frozen episode address and dynamic content. Parameter-free key masking and per-layer residual resets constrain the address channel within one shared multimodal trunk. e-architecturee-routing
- Source description
The paper combines benchmark success with a two-switch architectural ablation and address-swap intervention, testing both performance and whether behavior responds to the address interface. e-oa-ablatione-swap
Figure 2. One shared representation supplies future-state supervision and action generation. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the downward arrows from observations and instruction. The Qwen3VL branch supplies noun phrases for SAM3; DINOv3 contributes masked appearance features. The blue slot stream passes through its own adapter, while the other streams use the existing token embeddings. Sequence assembly ends with the starred action query. At the bottom, blue slot hiddens feed content and pose prediction; the pink query feeds the action head. The yellow image-VQ branch is an additional training objective. Appendix B clarifies that only third-person VQ codes enter the default sequence, although wrist observations also contribute to object slots. e-architecturee-preprocessinge-headse-inference
What it supports. The diagram supports a shared-backbone interpretation of joint world/action prediction. It does not show predicted future states feeding into the action head or a planning loop. Algorithm 1 instead returns both outputs; the action chunk is obtained with four flow-head Euler evaluations after the shared trunk pass.
Where the evidence stops. The world panel depicts objects, but its output is content and pose regression, not a rendered rollout. The VQ branch is disabled at inference. Simulator object poses and frozen perception remain substantive dependencies.
2. Motivation
2.1 The problem and the proposed response
A language-conditioned robot must keep identifying the same target when viewpoint or scene arrangement changes. The authors argue that holistic world latents mix identity with context, making target selection unstable even when objects remain visible. OA-WAM tests whether an explicit, persistent addressing interface helps under these shifts. e-problem
2.2 What this reading follows
Consider a robot asked to move one mug while another sits nearby. OA-WAM gives each detected object an episode-long address and a changing description of its current state. Its transformer restricts slot keys to an address channel, then predicts future object states and an action chunk from shared hidden representations. The useful question is whether this interface improves target binding when the camera or scene moves. The internal ablations support that hypothesis in simulation. They also expose a boundary: future prediction does not improve every reported metric, perception can still fail, and inconsistencies in aggregate scores and implementation details prevent a clean reproduction from the PDF alone. e-probleme-architecturee-routinge-oa-ablatione-world-ablatione-reporting-gaps
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 | One Model |
| Prediction paradigm | Joint prediction |
| Quadrant | Q1 · One Model × Joint prediction |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Q1 classification is supported by a common transformer sequence and hidden states feeding parallel world and action heads, not merely by joint loss training. Separate frozen perception modules do not create a second policy/world backbone. Future prediction is auxiliary at inference and action generation is flow matching, not inverse dynamics. The latent-prediction label fits; a specific JEPA architecture is not established. e-architecturee-headse-inference
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 an object's address from its changing state
Start at the beginning of an episode. Qwen3-VL interprets the instruction with the initial observation, and its noun phrases prompt SAM 3. DINOv3 supplies appearance information for the resulting masks. A learned address function combines the label with initial appearance, then caches that identity for the episode. Another function updates content as observations change. The same object can therefore move while retaining its address. The 320-dimensional slot also contains temporal and role information before projection into the trunk. This construction still needs correct initialization and tracking: two ambiguous candidates or a missed mask cannot be repaired simply by freezing a key. Appendix B's simulator-derived pose is another input dependency. The architecture makes an extracted object easier to address; it does not solve perception or prove end-to-end semantic grounding. e-architecturee-preprocessinge-adaptere-problem
Figure 3. Temporal visibility and address-only key projection are distinct constraints. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Q down the left edge and K across the top: a colored cell indicates an allowed query-to-key connection. The lower-triangular layout prevents later frame groups from leaking into earlier ones. Within each frame, the red slot block permits mutual slot attention. Action rows can read preceding visual and slot tokens, while same-frame slots cannot read those actions. The final yellow ACT_Q row pools accessible history. This is the visibility mask; Equation (3) adds a separate operation inside key projection, retaining only the first 32 residual coordinates at slot positions. The reset hook is specified in the accompanying text, not drawn here. e-routinge-adaptere-theory
What it supports. The architecture separates what a token may read from which part of a slot defines its key. Shared within-frame RoPE positions support permutation equivariance when masks and geometry are permuted consistently. That property concerns reordering the same slots; it is not a theorem that arbitrary visual perturbations leave actions unchanged.
Where the evidence stops. Full queries and the default geometry bias still affect attention logits. The source also leaves unclear how the unrestricted slot adapter establishes raw address coordinates before the first block. Address-only keys are narrower than context-independent routing.
5.2 Understand why the mask also needs a reset
A key mask alone selects coordinates, not their meaning. After a transformer block, residual updates could mix changing content into those selected coordinates. OA-WAM therefore restores the cached address after every block. This explains why the mask and reset are tested separately in the factorial: V3 keeps the mask but allows drift, whereas V1 keeps the reset but lets keys read all coordinates. The source's mechanistic claim is strongest at this tensor-level distinction. Reader analysis adds two qualifications. First, a fixed key does not fix an attention weight when queries or relative-geometry biases change. Second, the input adapter mixes all slot components, and the text does not specify an initial overwrite before the first block. A reproduction should verify that invariant directly instead of assuming the post-block hook establishes it retroactively. e-routinge-adaptere-oa-ablation
5.3 Use future prediction as supervision, then test executed behavior
The final slot hiddens predict next-frame content and pose, while the action query conditions a flow head over a whole action chunk. During training, Gaussian noise is interpolated toward demonstrated actions and the flow head learns the corresponding velocity. At inference, four Euler evaluations sample the chunk after one shared trunk pass; the auxiliary VQ head is disabled. No documented step feeds a generated future back into action selection. Table 5 tests whether world supervision improves the learned policy and finds a camera benefit, even though its printed aggregate does not improve. Table 7 tests the address operations while retaining the perception stack. Read these results as complementary: one probes the training objective, the other the routing interface. Neither substitutes for real-robot validation or an independently measured planning capability. e-headse-inferencee-world-ablatione-oa-ablatione-limitations
5.4 Training and inference
During training
Appendix G reports Stage 0: Chameleon-7B warm-start adaptation for about 600k steps on 2.5T web/robot tokens, with world/VQ losses. Stage I aligns slots/world prediction for 50k steps on LIBERO, DROID and an OXE subset. Stage II uses standard LIBERO only for 100k steps, training rank-32 LoRA plus adapters/heads, approximately 127M parameters. Base trunk weights stay frozen in this final route. e-training
Stage II combines action flow matching, object-world MSE, VQ prediction, distractor consistency and weak target/reference hints. World/VQ/consistency/role weights are 0.5/0.04/0.1/0.05; consistency warms during the first 30%, and role hints stop halfway. Detached original outputs supervise consistency after distractor permutation/insertion. Training features are cached; RGB augmentation is absent and LIBERO-Plus perturbations are held out. e-objectivee-preprocessing
During inference
Cache initial addresses and language parsing, refresh observations and content, then compute one trunk forward pass. Four Euler evaluations of the smaller flow head transform Gaussian noise into the action chunk. Algorithm 1 returns future states alongside actions; generated VQ tokens are disabled and predicted states are not fed into a planner. Subsequent observations provide feedback, but execution cadence is inconsistently documented. e-inferencee-headse-runtime-gaps
5.5 Implementation flow
- Build persistent slots
Qwen3-VL parses the instruction and initial observation into noun phrases/relations; SAM 3 tracks masks and DINOv3 supplies appearance features. These modules are frozen. An address MLP uses the label and initial feature once per episode; a content MLP refreshes state each frame. Capacity is one robot plus 16 object slots, with masked padding. e-preprocessinge-architecture
- Assemble a shared history
The default four-frame sequence combines BPE instruction, third-person VQ codes, slots, discretized state and past actions, ending in ACT_Q. Wrist imagery contributes through slots. The slot adapter maps 320-dimensional states to 4096-dimensional embeddings; future action targets are excluded from the input. e-architecturee-preprocessinge-adapter
- Constrain keys, preserve content
Each of 32 transformer layers masks slot-key inputs beyond the first 32 coordinates and resets that residual slice after the block. Full queries and values still carry task/state information. Slots share per-frame RoPE positions; relative geometry also biases slot-pair logits. Thus the verified constraint concerns key inputs, not complete context-independent attention. e-routinge-adaptere-theory
- Read world and action outputs
Slot hidden states feed content and pose MLPs with masked next-frame MSE supervision. ACT_Q conditions a flow-matching MLP over the entire action chunk. The shared trunk also supports an auxiliary VQ head. There is no documented search over predicted futures before selecting an action. e-headse-inference
6. Experiments & results
OA-WAM gives a shared world/action transformer persistent object addresses alongside changing content. Its strongest evidence is improved simulated camera-shift robustness when address masking and residual resets are enabled. Future-slot prediction is auxiliary supervision, not an inference-time planner. Several implementation and aggregate-reporting inconsistencies limit reproducibility.
6.1 Read the original evidence
Table 1. OA-WAM has the highest listed averages on these two standard benchmarks. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Separate the LIBERO columns on the left from the WidowX visual-matching tasks on the right. Each block has its own Avg column and should be read under its own evaluation protocol. The highlighted row reports OA-WAM; compare its LIBERO average with VLA-JEPA and its SimplerEnv average with CoWVLA. Appendix H specifies 100 episodes per LIBERO suite and 25 per SimplerEnv cell, each with three seeds for OA-WAM. Comparator rows mostly come from the cited papers. Dashes preserve unavailable results; F1-VLA's missing SimplerEnv average reflects a different original aggregation, so it should not be filled by assumption. e-standarde-evaluation
What it supports. OA-WAM reports 97.8% on LIBERO versus VLA-JEPA's 97.2%, and 79.3% on SimplerEnv versus CoWVLA's 76.0%. These are differences of 0.6 and 3.3 percentage points. Individual tasks remain uneven: the highest average does not mean OA-WAM wins every suite or task.
Where the evidence stops. These are simulated task-completion scores, including SimplerEnv's visual matching. Baseline training and uncertainty are not standardized by importing table rows; no supplied confidence intervals establish the significance of small differences.
Table 2. The camera and geometric-average gains are clearer than the aggregate claim. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with Camera, Robot and Layout, the three columns summarized by Geo Avg. OA-WAM reaches 80.5, 89.6 and 82.8, giving the displayed geometric average of 84.3. Camera improves over Cosmos-Policy, but X-VLA remains slightly higher on Robot and pi0.5 higher on Layout. Next inspect the photometric and language columns rather than assuming geometry gains transfer to them. Sensor Noise is especially weak relative to Cosmos-Policy. The final delta row compares each column with that column's prior best, not with one common baseline. All differences between success percentages should be read as percentage points. e-robustnesse-evaluatione-reporting-gaps
What it supports. The table supports an addressability advantage on camera shifts and on the geometric average among the listed methods: +4.7 and +4.8 points respectively. The noise score is 75.6 versus Cosmos-Policy's 92.7. The authors print overall scores of 83.9 for OA-WAM and 85.7 for pi0.5, but their aggregation remains unresolved.
Where the evidence stops. The displayed seven axes average 86.6 for OA-WAM and approximately 86.49 for pi0.5, not the printed 83.9 and 85.7. No alternative weighting is specified. Preserve the original cells, but do not treat the printed aggregate ranking as verified arithmetic.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO standard manipulation Four simulation suites; standard demonstrations; 100 evaluation episodes per suite and three seeds. | 97.8; Spatial/Object/Goal/Long: 98.9/99.0/97.4/95.9. Mean task success (%) | VLA-JEPA: 97.2 average; reported advantage 0.6 percentage points. Strong in-distribution performance; imported comparator results and absent intervals limit claims about small differences. e-standarde-evaluation |
| SimplerEnv WidowX Bridge visual matching Four simulated Bridge tasks; official visual-matching protocol, 25 episodes per cell and three seeds. | 79.3; Spoon-Towel/Carrot-Plate/Stack-Cube/Eggplant-Basket: 83.0/71.1/65.0/98.2. Mean success (%) | CoWVLA: 76.0 average, a 3.3-point difference. Simulation with real-scene visual matching, not executed physical-robot success. Missing baseline cells remain unreported. e-standarde-evaluation |
| LIBERO-Plus geometric robustness Zero-shot perturbations after standard-LIBERO finetuning; official generator, three seeds. | Camera 80.5; robot 89.6; layout 82.8; Geo Avg 84.3. Success (%) and arithmetic mean across Camera/Robot/Layout | Camera exceeds Cosmos-Policy 75.8 by 4.7 points; Geo Avg exceeds pi0.5 79.5 by 4.8. Robot trails X-VLA 89.7; layout trails pi0.5 85.7. The table supports best camera and geometric average among listed rows, not leadership on every geometric axis. e-robustnesse-evaluation |
| LIBERO-Plus reported aggregate and sensor noise Same zero-shot seven-axis table. | Printed Avg 83.9; sensor noise 75.6. Reported success (%) | Printed pi0.5 Avg 85.7; Cosmos-Policy noise 92.7. Preserved as reported: the seven displayed axes average 86.6 for OA-WAM and about 86.49 for pi0.5, inconsistent with the printed aggregate. The claimed aggregate ranking cannot be independently reconstructed. e-robustnesse-reporting-gaps |
| Object-addressability factorial and swap binding Shared perception/tokenization/training setup; mask/reset combinations V2 off/off, V1 off/on, V3 on/off, V0 on/on. | Camera: 60.5/67.2/70.8/80.5; binding: 0.06/0.19/0.32/0.87. LP-camera success (%) and address-swap cosine | Full versus no OA adds 20.0 camera points, versus 2.4 on standard LIBERO. Holistic swap baselines score at most 0.09. The internal factorial supports a routing contribution. The external swap comparison lacks a fully specified equivalent intervention for holistic representations. e-oa-ablatione-swap |
| World objective and distractor consistency ablations Appendix I, supporting ablations on standard LIBERO and LIBERO-Plus. | World prediction raises camera 73.4→80.5 and LIBERO 95.6→97.8; printed LP Avg changes 84.5→83.9. Consistency raises layout 78.5→82.8, lowers permutation KL 0.21→0.04 and insertion drift 0.19→0.05. Success (%) and diagnostic drift | World-loss removal and joint removal of both consistency augmentations. The world objective helps camera success without improving the printed aggregate. 'Action only' still retains VQ supervision; claimed statistical flatness is not backed by disclosed intervals. e-world-ablatione-consistency |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 6. The two routing operations matter beyond supplying object slots alone. Original paper, p. 25 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the variant labels before comparing bars. V2 has neither mask nor reset; V1 keeps only the reset; V3 keeps only the mask; V0 enables both. All retain the paper's slot/perception setup. In panel (a), the medium-blue camera bars separate much more than the light-blue standard-LIBERO bars. The curved arrow points from V2's camera result to V0's and labels a 20.0-point increase; it is an annotation, not a fitted trajectory. Panel (b) uses a different metric, the cosine of intervention-induced motion toward the swapped target. The dotted line marks the largest reported holistic-baseline cosine. e-oa-ablatione-swape-training
What it supports. Camera success rises from 60.5 without OA to 80.5 with both operations, while standard LIBERO changes from 95.4 to 97.8. Swap binding rises from 0.06 to 0.87. The two partial variants sit between those endpoints, supporting the usefulness of both mask and reset within this controlled design space.
Where the evidence stops. The success axis begins at 50%, so bar heights exaggerate proportional differences. Binding measures response to an artificial address swap, not successful manipulation. Shared OA-pretrained weights and an unspecified equivalent holistic intervention limit broader causal conclusions.
Table 5. Future-state supervision helps the camera axis without improving the printed aggregate. Original paper, p. 24 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read this as an ablation of training supervision. The upper row removes the slot-world objective; the lower row retains it. The three columns measure standard LIBERO, the LIBERO-Plus camera axis and the printed LIBERO-Plus aggregate. The crop excludes the long caption, whose crucial qualification is that 'Action only' still keeps auxiliary VQ prediction. It therefore does not isolate all future prediction from action learning. The main method defines next-frame slot regression, although this caption calls the removed loss multi-step. That horizon mismatch remains unresolved, so the tutorial follows the explicit next-frame equations. e-world-ablatione-headse-reporting-gaps
What it supports. With world supervision, standard LIBERO improves by 2.2 points and camera success by 7.1 points. The printed aggregate instead decreases by 0.6 points. This supports a specific camera-robustness contribution; it does not establish a universal benefit from adding a world head or a need for inference-time future imagination.
Where the evidence stops. The caption calls the aggregate difference statistically flat without providing intervals. Table 2 already leaves aggregate computation unresolved. No independent world-prediction quality metric here explains whether improved content prediction, pose prediction or shared representation learning causes the camera gain.
7. Analysis & limitations
7.1 What the evidence leaves open
Validation is simulator-only and uses privileged object poses. Small, reflective, transparent, occluded or blurred objects can defeat extraction; distractor invariance assumes weak physical coupling. The failure audit identifies dynamics, perception, engineering and ambiguous/blocked tasks rather than one universal cause. e-limitationse-preprocessing
The sensor-noise deficit is attributed to perception, but no oracle-perception repair establishes that attribution. Appendix I mentions depth-based pose failures despite Appendix B specifying simulator poses. Assigning all holistic failures to 'policy' by convention is not a fair component-level causal comparison. e-limitationse-pose-gape-robustness
Key masking does not freeze queries, values or geometry-biased logits. Moreover, the unrestricted input adapter is not shown preserving address coordinates before the first block. Sixteen-slot capacity is not reconciled with larger active-object counts. These gaps weaken an unconditional architectural guarantee. e-routinge-adapter
No per-seed scores or confidence intervals establish significance. Table 2's aggregate arithmetic is unresolved; Table 8's partial failure accounting should not be read as total error. Several baseline rows lack specific Appendix-H attribution, and imported rows do not constitute matched retraining. e-reporting-gapse-evaluation
7.2 Questions for discussion
- Would the mask/reset advantage persist with independently trained controls rather than a shared OA-pretrained initialization?
- Does repairing segmentation or pose under sensor corruption close the measured noise gap?
- Which aggregation weights, execution cadence and first-layer address invariant produced the reported results?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the claimed Stage-0 checkpoint, exact feature caches, quantization statistics, dataset splits and benchmark configurations. Stages I/II report eight A100-80GB GPUs for 3–4 days each. Stage 0 reports 384 GPUs and about 166k A100-hours, but its twelve-node description conflicts with that GPU count; Stage-II microbatch arithmetic also conflicts with the stated global batch. e-traininge-training-gaps
Resolve when slot/world modules enter Stage 0, first-layer address initialization, pose-loss weight, future-content target gradients, world-loss normalization, possible double VQ weighting and optional role readout. Detailed dual-view/Euler tables are absent; rotation comparisons are explicitly unreported. A runnable configuration cannot be inferred safely from these descriptions alone. e-adaptere-training-gapse-implementation-gaps
Measure perception, trunk and Euler latency separately before reproducing control claims: the conclusion's 5.6-ms trunk/head figure conflicts with Appendix G's 80+10 ms. The 233-ms total includes first-only parsing, and 4.3 Hz is not reconciled with 16-step execution at 20 Hz. The edition proposes a controlled routing factorial and an oracle-perception noise test; neither was run. e-runtime-gapse-inferencee-oa-ablatione-pose-gap
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: Separate training adaptation from test-time routing damage
Reader-proposed check, not performed: starting from the same verified Stage-0 checkpoint, train all four mask/reset configurations with identical Stage-II data, seeds, budgets, role losses and cached perception. Also toggle each switch only at evaluation in the fully trained V0 model. Pair initial scenes and action-sampling noise across conditions, and measure camera-angle success, standard LIBERO success, swap-binding cosine and first-32-coordinate drift at every layer, including layer zero. A persistent camera advantage after matched retraining would support the routing bias; a gap confined to abrupt evaluation toggles would suggest distribution-shift damage. Record the shared checkpoint's OA pretraining as a remaining confound rather than claiming a from-scratch architecture comparison. e-traininge-adaptere-routinge-oa-ablatione-evaluation
Check 2: Test whether perception repair closes the sensor-noise gap
Reader-proposed check, not performed: hold a trained policy, initial states, instructions and action-noise seeds fixed across clean and sensor-corrupted observations. Compare ordinary perception, oracle object masks with estimated poses, oracle poses with ordinary masks, and both oracle components together. Cache the same initial addresses before corruption; separately test corruption at episode initialization. Report per-axis task success, target tracking errors and denominators, and time perception separately so latency does not masquerade as binding failure. If oracle repair restores most lost success, the perception explanation gains support; persistent failures would implicate the remaining VQ/content pathway or policy. First resolve whether the published evaluation actually used simulator or estimated poses. e-preprocessinge-robustnesse-limitationse-pose-gape-runtime-gaps
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version block; Figures 1–9; Tables 1–8; Algorithms 1–2; method equations; and appendix preprocessing, architecture, objectives, hardware, evaluation and failure details. Every final crop was viewed, including the revised architecture crop. Figure 2's branch directions and Figure 3's Q/K orientation and slot blocks were checked against the equations and Appendix C; Figure 6's arrow was checked against Table 7. All claim-supporting non-reference pages are included. Reference pages 10–15 were read as text only. No separate supplement, executable implementation or external baseline source was inspected; aggregate, latency, pose-source and training-description inconsistencies remain explicit.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28. 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
- Title and author block; arXiv:2605.06481v1, 7 May 2026
- Abstract
- 1 Introduction
- 2 Related Works
- 3 Method, Sections 3.1–3.5
- 4 Experiments, Sections 4.1–4.3
- 5 Conclusion and limitations
- References
- A Pseudocode, Algorithms 1–2
- B Six-path tokenization and data preprocessing
- C Slot adapter, sequence template, and slot-aware trunk
- D Prediction-head architectures
- E Equivariance property
- F Loss components and augmentations
- G Three-stage training and inference latency
- H Evaluation protocol
- I Ablations, fairness controls, failure audit and benchmark galleries
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The complete supplied 28-page text was read in all ten chunks. Only the supplied v1 was inspected; no later revision or cross-version comparison was available. Title and all ten authors match the catalog after name-order normalization.
- 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 all nine figures, all eight numbered tables and the relevant appendix pages.
- Separate supplemental material availability has not been fully verified.
- No separate supplement, code, checkpoint, dataset or cited paper was supplied for independent inspection. No experiments were reproduced. Source claims of released scripts/checkpoints remain unverified.
- Reference pages 10–15 were read as text, not rendered. All other PDF pages were rendered and visually inspected.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title, author block and arXiv margin stamp
The title matches the catalog. Authors in order are Yushan Liu, Peibo Sun, Shoujie Li, Yifan Xie, Lingfeng Zhang, Xintao Chao, Shiyuan Dong, Fang Chen, Xiao-Ping Zhang and Wenbo Ding. Affiliations are Tsinghua University, Shanghai Jiao Tong University and Nanyang Technological University. The artifact is marked arXiv:2605.06481v1 [cs.RO], 7 May 2026, Preprint.
Go to primary source ↓e-problemPDF pp. 1–4, Abstract and Section 1; Figure 1
The authors attribute manipulation failures under scene shifts to holistic representations entangling target identity with context. They propose persistent object addresses while explicitly conditioning their grounding claim on successful upstream extraction.
Go to primary source ↓e-architecturePDF p. 3, Figure 2 and caption; p. 5, Sections 3.1–3.2, Equations (1)–(2)
A shared Chameleon-style trunk processes multimodal history with object slots and a trailing action query. Each 320-dimensional slot contains 32 address, 256 content, 16 temporal and 16 role dimensions. Separate heads predict future object states and a 16-by-7 action chunk.
Go to primary source ↓e-preprocessingPDF pp. 17–18, Appendix B, preprocessing and Paths T1/T2/I-A/I-B/S/A-d
Four historical frames are used. Images are resized to 256 by 256 without RGB augmentation; only third-person VQ codes enter the default sequence, while wrist views contribute to slots. Frozen Qwen3-VL-4B parsing, SAM 3 masks and DINOv3 ViT-L/16 features support slots; pose uses simulator ground truth. Masks are deduplicated at IoU >0.5 and feature cosine >0.9. The first six action dimensions use training-split 1st/99th percentile clipping and scaling to [-1,1], with binary gripper commands.
Go to primary source ↓e-adapterPDF pp. 18–20, Appendix C, slot adapter, injection, reset and address capacity
The slot adapter is an unrestricted 320-to-4096-to-4096 MLP. A post-block hook writes cached addresses into the first 32 residual coordinates and stops gradients there; the input embedding still trains the address MLP. The text specifies 16 object slots but later describes up to 24 objects and 27 active addresses without an overflow policy. An explicit pre-first-block address overwrite is not given.
Go to primary source ↓e-routingPDF p. 5, Figure 3 and Equation (3); pp. 19–20, Appendix C, attention organization
Slot keys zero residual coordinates beyond the first 32 before the existing key projection. Queries and values retain full inputs. Slot attention is bidirectional within each frame, frame history is causal, action tokens read same-frame visual/slot tokens but not conversely, and ACT_Q reads prior positions. Shared slot RoPE positions and a default relative-SE(3) slot-pair logit bias are specified.
Go to primary source ↓e-headsPDF p. 6, Section 3.4, Equations (4)–(5); p. 20, Appendix D
The world head regresses next-frame 256-dimensional content and 9-dimensional pose, masking robot and padding slots. The flow head conditions on ACT_Q, learns the velocity from Gaussian noise toward demonstrated 16-step actions, and samples using four Euler steps. The VQ head is auxiliary and disabled at inference.
Go to primary source ↓e-objectivePDF p. 6, Section 3.5, Equation (6); p. 16, Algorithm 2; pp. 21–22, Appendix F
Training combines action flow matching, world MSE, VQ cross-entropy, distractor consistency and weak role hints. Weights are 1, 0.5, 0.04, 0.1 and 0.05, with consistency warmed over the first 30% and role hints removed after half of Stage II. Consistency compares original detached outputs against slot-permuted or inserted-distractor outputs using attention KL and action L2.
Go to primary source ↓e-trainingPDF pp. 22–23, Appendix G, Stages 0/I/II and hyperparameter table; p. 21, module accounting
Appendix G reports approximately 600k Stage-0 steps and 2.5T web/robot tokens after Chameleon-7B initialization; Stage I uses 50k world-loss alignment steps on LIBERO, DROID and an OXE subset; Stage II uses 100k steps on standard LIBERO only with rank-32 LoRA, adapters and heads, about 127M trainable parameters. Stages I/II each report 3–4 days on eight A100-80GB GPUs.
Go to primary source ↓e-inferencePDF p. 16, Algorithm 1; p. 20, Appendix D, Equation (7); p. 23, Appendix G, latency breakdown
Inference caches initial addresses and language parsing, refreshes perception/history, runs one trunk forward pass and four action-MLP Euler evaluations. World output is returned alongside the action chunk. Appendix G says chunks are processed every 16 simulator steps and reports 233 ms per chunk, including first-chunk parsing.
Go to primary source ↓e-evaluationPDF p. 6, Section 4.1; p. 23, evaluation runtime; p. 24, Appendix H
All evaluations are simulation. LIBERO uses four suites, 100 episodes and three seeds per suite. LIBERO-Plus uses the official perturbation generator with no perturbation training; p. 23 specifies 100 episodes per axis per seed. SimplerEnv uses WidowX Bridge visual matching, 25 episodes per cell and three seeds. Most comparator rows are imported; LIBERO-Plus pi0.5 is the authors' checkpoint evaluation. Per-seed values and confidence intervals are not supplied.
Go to primary source ↓e-standardPDF p. 7, Table 1, OA-WAM and comparator rows
OA-WAM reports LIBERO Spatial/Object/Goal/Long scores 98.9/99.0/97.4/95.9 and average 97.8%, versus VLA-JEPA 97.2%. SimplerEnv Spoon-Towel/Carrot-Plate/Stack-Cube/Eggplant-Basket scores are 83.0/71.1/65.0/98.2 and average 79.3%, versus CoWVLA 76.0%. Missing cells are dashes, not zero.
Go to primary source ↓e-robustnessPDF p. 8, Table 2, OA-WAM, pi0.5, X-VLA and Cosmos-Policy rows and delta row
OA-WAM reports camera 80.5, robot 89.6, layout 82.8, Geo Avg 84.3, light 96.5, background 95.9, language 85.3, noise 75.6 and Avg 83.9%. Comparators include Cosmos-Policy camera 75.8/noise 92.7, X-VLA robot 89.7, and pi0.5 layout 85.7/Geo Avg 79.5/Avg 85.7.
Go to primary source ↓e-oa-ablationPDF p. 9, Table 3; p. 25, Table 7 and Figure 6
V2 (mask off/reset off), V1 (off/on), V3 (on/off) and V0 (on/on) report LIBERO 95.4/96.3/96.6/97.8, LP camera 60.5/67.2/70.8/80.5 and swap binding 0.06/0.19/0.32/0.87. The paper describes a shared perception stack, tokenization, trunk weights and training pipeline. Figure 6's arrow labels the V2-to-V0 camera change as +20.0 percentage points.
Go to primary source ↓e-swapPDF p. 9, Section 4.3 A2 and Table 4; p. 8, Figure 5(c); p. 26, Table 8 caption
The authors swap the target address with another in-scene slot address, hold other inputs fixed, and measure cosine alignment of the residual end-effector trajectory with displacement toward the swapped target. OA-WAM scores 0.87, V1 0.19, mean-pool head 0.18 and eight holistic baselines 0.04–0.09. An equivalent address intervention for models without an address subspace is not operationally specified.
Go to primary source ↓e-world-ablationPDF p. 24, Appendix I, Table 5 and caption
Removing the world loss gives LIBERO 95.6, LP camera 73.4 and LP Avg 84.5, versus 97.8/80.5/83.9 with world prediction. The caption calls the aggregate difference statistically flat but supplies no intervals. Its action-only configuration retains auxiliary VQ prediction, and it describes the removed world loss as multi-step despite the next-frame formulation.
Go to primary source ↓e-consistencyPDF p. 24, Table 6; p. 21, Appendix F
Without distractor consistency, LP layout is 78.5%, permutation KL 0.21 and insertion drift 0.19; full loss gives 82.8%, 0.04 and 0.05. Both permutation and insertion components are removed together.
Go to primary source ↓e-theoryPDF p. 19, frame-internal positional sharing; p. 21, Appendix E
The equivariance proposition requires shared per-frame RoPE position, shared slotwise operations and consistent permutation of geometry and padding while fixing the robot slot. It concerns slot order, not guaranteed invariance to arbitrary scene changes.
Go to primary source ↓e-limitationsPDF p. 9, Conclusion limitations; p. 26, Appendix I, failure-mode breakdown
The source limits validation to simulation and notes reflective, transparent, small, occluded and blurred objects, plus weak target/distractor coupling. Its post-hoc 300-failure sample contains 121 action/dynamics, 89 perception, 47 engineering and 43 out-of-scope failures. Table 8 assigns all holistic failures to policy by convention because those models expose no perception cache.
Go to primary source ↓e-reporting-gapsPDF p. 8, Table 2 and caption; p. 24, Appendix H; p. 26, Table 8 and failure discussion
The seven displayed OA-WAM axis values sum to 606.2 and the pi0.5 values sum to 605.4; these do not yield the printed Avg values 83.9 and 85.7 under a simple mean. No alternative weighting is specified. Table 8's OA-WAM perception/policy rates 3.4/7.2 omit other failure categories and cannot be treated as the total failure rate. Appendix H does not itemize all Table 2 comparator attributions.
Go to primary source ↓e-runtime-gapsPDF p. 9, limitations; p. 23, Appendix G, inference breakdown
The conclusion gives approximately 5.6 ms for trunk plus head, while Appendix G gives 80 ms trunk and 10 ms action head. The 233-ms total includes a 43-ms parse explicitly cached after the first chunk. Its claimed 4.3-Hz control rate is not reconciled with execution every 16 steps in a 20-Hz simulator.
Go to primary source ↓e-training-gapsPDF p. 19, backbone specs; p. 21, module table; pp. 22–23, Appendix G
Appendix C places new modules/reset after pretraining, whereas Appendix G uses slot-world loss in Stage 0. The reset is nonparametric in p. 21 but called trainable in Stage I. Stage 0 lists 384 GPUs but twelve eight-GPU nodes; Stage II lists global batch 128 but eight GPUs times microbatch four times accumulation eight. These descriptions are internally inconsistent.
Go to primary source ↓e-implementation-gapsPDF p. 6, Equations (4)–(6); p. 16, Algorithm 2; pp. 18–23, Appendices B–G; pp. 25–26, omitted comparisons
The value of pose-loss weight lambda_p and treatment of learned future-content targets are not specified. World-loss averaging differs between Equation (4) and Algorithm 2. VQ weighting is described both inside cross-entropy and outside the loss. Detailed dual-view/Euler ablation tables promised in B/D are absent from G; rotation alternatives are explicitly unreported. Optional role attention and four-role readout descriptions do not fully define their connection to the ACT_Q-conditioned flow MLP.
Go to primary source ↓e-pose-gapPDF p. 17, state preprocessing; p. 18, Path I-B; p. 26, failure-mode breakdown
Appendix B describes simulator-ground-truth object pose and places depth-assisted deployment outside the main empirical claim. Appendix I nevertheless attributes some evaluated failures to depth-based pose. The pose source for that failure audit is not reconciled.
Go to primary source ↓e-galleryPDF pp. 27–28, Figures 7–9 and captions
Task galleries illustrate the four LIBERO suites and four WidowX Bridge tasks. The perturbation gallery distinguishes layout, background, light, camera, robot initialization, language and several sensor-corruption examples; it is illustrative rather than an additional measured experiment.
Go to primary source ↓8.5 Primary sources
OA-WAM: Object-Addressable World Action Model for Robust Robot Manipulation ↗
PDF · 15,176 extracted words
Source fingerprint
83df65b170d50640dbaf0e23dad5c37f324ef22e26dfc022167baedc4790ca7d