AIM: Intent-Aware Unified world action Modeling with Spatial Value Maps
1. Paper overview
In one sentence: AIM uses predicted contact maps to connect visual foresight to robot actions, trading explicit simulator supervision for an interpretable spatial interface whose causal benefit remains incompletely tested. e-motivatione-overviewe-maske-rle-aggregatee-figure-conflicte-protocol
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that predicting scene appearance leaves manipulation intent implicit: a controller still needs to identify where useful contact should occur. AIM inserts an aligned spatial representation between visual foresight and action generation to concentrate this information. Whether that interface itself causes the performance gains requires controls beyond the reported baseline comparisons. e-motivatione-overviewe-protocol |
| Core mechanism | Source description A video-prior model predicts future RGB and contact-oriented value maps, coupled to a narrower action transformer through shared masked attention. The map is used during action generation, rather than serving only as an auxiliary training target. e-architecturee-mask |
| A key reported result | RoboTwin 2.0 aggregate manipulation success: AIM: Easy 94.0%, Hard 92.1%; Table 2 combined Average 93.1%. Reported average success rate (SR; higher is better). 50 simulation tasks, Easy/Hard; 30K training trajectories. Section 6 states common data, task definitions and success criteria; split details are absent. Table 2: LingBot-VA 92.9%/91.6%, Fast-WAM 91.9%/91.8%, Motus 88.7%/87.0%, π0.5 82.7%/76.8% (Easy/Hard). AIM leads the reported aggregate comparison. Differences from Motus are 5.3/5.1 percentage points using Table 2's rounded entries; the prose says +5.0 for Hard. No uncertainty or significance test is supplied. e-protocole-aggregate |
| Reading caution | Reader analysis Figure 2 draws future-frame-to-action and action-to-map arrows, conflicting with the future-RGB → map → action dependence in Equation (10). Its 'PPO Post-Training' label also conflicts with its GRPO caption and Section 4.3. The equations/text specify the intended mechanism; the diagram is not a reliable wiring specification. e-figure-conflicte-maske-rl |
Core contributions
- Source description
A video-prior model predicts future RGB and contact-oriented value maps, coupled to a narrower action transformer through shared masked attention. The map is used during action generation, rather than serving only as an auxiliary training target. e-architecturee-mask
Figure 1. AIM exposes a spatial interface between predicting a future and choosing an action. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the output blocks above the shared model: the left schematic has future video and action, while AIM adds future value and a self-distillation connection. The lower video, instruction and state blocks provide a conceptual inventory of inputs. To understand the added branch precisely, read Equation (1) on the same page: future observations and maps are modeled together, and the action distribution conditions on those maps and history. Section 4.2 then gives the attention restrictions implementing this dependence. The figure summarizes the idea; the factorization and visibility sets explain how future information can reach the action head. e-overviewe-tokenizatione-maske-rl
What it supports. The map pathway participates in control as well as training. AIM makes future interaction regions an explicit predicted representation, and later uses their responses as a dense reward for the action head. This provides an inspectable intermediate output, although interpretability alone does not establish that it improves execution.
Where the evidence stops. The 'State' and 'Prop. Tokenizer' blocks are schematic: the formal prefix specifies observations and recent actions without fully defining a separate state-vector interface. The self-distillation arrow summarizes post-training, not an extra deployment optimizer.
2. Motivation
2.1 The problem and the proposed response
The authors argue that predicting scene appearance leaves manipulation intent implicit: a controller still needs to identify where useful contact should occur. AIM inserts an aligned spatial representation between visual foresight and action generation to concentrate this information. Whether that interface itself causes the performance gains requires controls beyond the reported baseline comparisons. e-motivatione-overviewe-protocol
2.2 What this reading follows
A useful imagined future must tell a robot where to interact. AIM makes that requirement explicit by predicting spatial value maps alongside future RGB frames, then restricting the action head's future information to the map pathway. A second training stage turns the frozen map predictor into a reward source for improving the action policy. This reading follows the interface, the actual visibility rules and the difference between supervised and post-trained results. The reported RoboTwin success rates are strong, but diagram/text conflicts, missing mechanism controls and limited evaluation detail make the evidence less decisive about why the method works. e-motivatione-overviewe-maske-rle-aggregatee-figure-conflicte-protocol
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 | Not assigned |
| Architecture | Not assigned |
| Prediction paradigm | Not assigned |
| Quadrant | Not assigned |
This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.
3.1 Evidence-based assessment
Insufficient evidence to decide
The recorded catalog fields are all 'Not assigned', so no existing quadrant can be confirmed. Architecture evidence supports a unified One Model interpretation at the coupled-network level: specialized branches share masked attention during generation. Its prediction mechanism combines joint future RGB/map/action denoising with value-conditioned inverse-dynamics action prediction and later action-only RL. This assessment does not assign or overwrite a catalog quadrant. e-overviewe-architecturee-maske-rl
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 contact structure before using it as intent
AIM's map supervision begins with simulator geometry. For a pick, successful grasp contact supplies vertices that are projected into each camera and smoothed. For a place, stabilized object/support contact supplies the target region. Packing these maps exactly like the RGB views and using the same VAE gives corresponding spatial layouts. My interpretation is that the map is a learned contact-oriented interface: it tells the action branch which image regions deserve attention as the task progresses. It should not automatically be read as a calibrated probability of success or an expected-return value function, because the supplied annotation procedure defines contact heat maps and gives no Bellman target. This distinction matters when the same predicted map later provides RL reward: useful contact supervision does not guarantee accurate reward values for every action the policy may explore. e-datae-tokenizatione-overviewe-rl
Figure 2. Joint supervised prediction supplies a frozen spatial teacher for action-only post-training. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the camera and language encoders, then locate the Wan2.2 mixture-of-transformers block. Stage I supervises future frames, maps and actions; the lower panel converts predicted actions and map responses into sparse and dense feedback. For the actual information flow, follow Equation (10) on page 6: the value stream can read future RGB, and the action stream can read future values while direct future-RGB access is masked. Section 4.3 freezes video/value prediction and updates only actions. The drawn frame-to-action and action-to-map arrows conflict with that formal dependence, so they cannot be used as a literal implementation guide. e-tokenizatione-architecturee-losse-maske-rle-figure-conflict
What it supports. The intended design separates learning a spatial predictor from adapting the action policy through environmental feedback. The frozen map contributes dense reward at projected action targets, alongside sparse task signals. This is a specific teacher–policy relationship; it does not mean the map predictor is continually corrected during RL.
Where the evidence stops. The unchanged graphic says 'PPO Post-Training'; its caption and Section 4.3 say GRPO. Its direct frame-to-action arrow also contradicts Equation (10). This edition explains the verified equations and prose, while leaving the source's diagram and optimizer discrepancy unresolved.
5.2 Read the mask as the operational definition of the bridge
The three future streams are denoised within a coupled network, but they do not have symmetric access. In Equation (10), future video can read observations, historical actions and language; future values can read observations and future video; actions can read the current observation, historical actions and future values. Each stream also sees its own future tokens. This gives the action head a route to predicted spatial structure while excluding a direct future-RGB route. Language enters the video model, so its influence on future-conditioned actions is mediated by the world/value representations. The branch-specific projections and feed-forward layers mean that shared attention does not imply identical parameters everywhere. Figure 2 is misleading about key arrows; the explicit visibility sets, together with the text, are the source's clearest specification of the intended dependence. e-architecturee-maske-figure-conflict
5.3 Separate policy refinement from evidence for the spatial mechanism
Stage1 learns from demonstrations and contact-map labels. During post-training, the action policy interacts with the simulator while the video and value predictors stay frozen. A candidate target earns dense feedback by projecting into the predicted map, and sparse feedback records task success or completion. The paper calls the clipped-ratio update GRPO, despite the schematic's PPO label. My interpretation is that this creates a fixed learned teacher for the action policy, with a possible failure mode if erroneous high-value regions become exploitable. Table 2 tests the whole post-training stage: its reported gains are 1.0 point on Easy and 0.1 on Hard, with a displayed-row aggregation ambiguity for Hard. It does not show that the dense component is necessary, that maps are calibrated, or that intent-causal attention causes the supervised advantage. e-losse-rle-figure-conflicte-stagee-row-arithmetice-protocol
5.4 Training and inference
During training
Pick labels project effective grasp-contact vertices through calibrated cameras and apply Gaussian smoothing adapted to camera parameters and depth. Place labels project object/support contact after stabilization under a small center-of-mass velocity threshold. These supervise interaction heat maps; the paper supplies no Bellman-return target for them. e-data
Stage1 minimizes weighted RGB flow-matching, map flow-matching and inverse-dynamics action losses. RL starts from Stage1, freezes the video generator and value head, and updates only the action head using the combined dense/sparse reward. Section 4.3 calls the clipped-ratio update GRPO. e-losse-rl
During inference
Initialize future RGB, map and action tokens from noise and denoise the coupled streams under the mask. Decode a world/action chunk and roll out autoregressively with a KV cache; newly appended real observations and predicted tokens receive fresh computation. The distribution factorization describes information dependence, not a separately documented serial sampling schedule. e-overviewe-architecturee-mask
Evaluation measures simulated task execution. Future-frame quality alone does not establish control success. The source does not specify how many actions are executed before feedback, denoising-step counts or measured control latency. e-architecturee-protocol
5.5 Implementation flow
- Encode aligned observations and maps
Pack head and wrist views into a T-pose canvas. Encode RGB and identically packed maps with the pretrained Wan2.2 VAE; project continuous actions through a lightweight MLP. T5 instruction features condition the video model through cross-attention. The text describes a black map input at initialization, separately from future-token Gaussian-noise initialization. e-tokenizatione-architecture
- Couple specialized streams
Initialize from Wan2.2-TI2V-5B. The action head has the same depth and smaller width. Streams use their own Q/K/V projections, a common attention dimension, shared masked attention, and separate feed-forward transformations. RGB and map tokens follow a joint flow-matching trajectory; the value stream also receives a learned value-noise token. e-overviewe-architecture
- Restrict future access
Equation (10) lets future RGB attend current/past observations, past actions, language and its own stream. Value tokens see current/past observations, future RGB and their own stream. Action tokens see current observations, past actions, future maps and their own stream. Direct future-RGB and direct language access are absent from the action visibility set. e-mask
6. Experiments & results
AIM jointly generates future RGB observations, spatial contact maps and continuous robot actions. A masked mixture-of-transformers architecture makes the predicted maps the action head's only route to future information. Supervised learning uses simulator-derived contact labels; post-training freezes world/value prediction and refines actions using map responses plus task rewards. The paper reports 94.0% Easy and 92.1% Hard success on RoboTwin 2.0, but the small post-training improvement, missing mechanism ablations and inconsistent schematic require careful interpretation.
The complete PDF contains three figures and two tables, all included here. It has no dedicated ablation figure or separate map-, mask-, or reward-component ablation; Table 2's Stage1/AIM comparison is the available stage-level diagnostic. Figure 3 provides execution snapshots without map overlays or projected targets, so the claimed spatial alignment has no directly inspectable diagnostic visual. No quantitative video/map-quality evaluation or physical-robot experiment is supplied. e-protocole-stagee-qualitativee-data
6.1 Read the original evidence
Table 1. Task-level results reveal both strong baseline gains and persistent failures. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each pair of columns as Easy then Hard, and compare the same row across methods. Stage1 and AIM are the most useful adjacent pair for examining the RL stage; Motus is a useful external comparison. Place Mouse Pad reaches 97%/95% for AIM against 66%/68% for Motus, yet Stage1 already has 97%/96%. Blocks Ranking Size reverses the favorable picture: AIM scores 47%/43%, below both Motus and X-VLA. Use the bottom row for the reported aggregate, but inspect individual rows before interpreting the system as uniformly reliable. Colored best-cell highlights are not uncertainty estimates. e-task-contraste-protocole-row-arithmetice-stage
What it supports. The strongest task gains coexist with clear failures. Scan Object reaches 100%/98% and Turn Switch 100%/98%, while Hanging Mug remains at 43%/42%. These rows support a heterogeneous performance profile, rather than a claim that spatial maps solve every manipulation skill.
Where the evidence stops. Trial counts and confidence intervals are absent. Reader averaging of the displayed Stage1 Hard cells gives 92.1%, while the printed average is 92.0%; AIM gives 92.1% by both routes. Unreported rounding or weighting could matter for that small difference.
Figure 3. Execution snapshots show selected manipulation stages under Easy and Hard scene conditions. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each row horizontally in the direction of the task-process arrow. The left and right halves show Easy and Hard examples of the same named task, rather than two competing policies. The Hard panels visibly contain more varied backgrounds and distractor objects. Track the gripper and manipulated object across the displayed snapshots, then return to Table 1 for measured task success. For example, the visually illustrated press-stapler task has AIM success of 96%/94%. The strips contain no temporal scale or predicted-versus-observed labels, so they provide selected execution context rather than a quantitative test of video fidelity or control speed. e-qualitativee-task-contraste-protocol
What it supports. The examples make the evaluated manipulation settings concrete and show the authors' selected execution progressions. They can help explain what a task involves, but the success rates come from the separate evaluation tables, not from counting completed-looking snapshot sequences.
Where the evidence stops. No value-map overlay, action-target marker or quantitative spatial alignment diagnostic is visible. The adjoining text's claim that maps capture meaningful interaction regions and rule out shortcuts cannot be verified from these strips. Selected stills also do not establish continuous execution quality.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0 aggregate manipulation success 50 simulation tasks, Easy/Hard; 30K training trajectories. Section 6 states common data, task definitions and success criteria; split details are absent. | AIM: Easy 94.0%, Hard 92.1%; Table 2 combined Average 93.1%. Reported average success rate (SR; higher is better) | Table 2: LingBot-VA 92.9%/91.6%, Fast-WAM 91.9%/91.8%, Motus 88.7%/87.0%, π0.5 82.7%/76.8% (Easy/Hard). AIM leads the reported aggregate comparison. Differences from Motus are 5.3/5.1 percentage points using Table 2's rounded entries; the prose says +5.0 for Hard. No uncertainty or significance test is supplied. e-protocole-aggregate |
| Effect of RL post-training Stage1 supervised checkpoint versus final AIM; same Easy/Hard simulation benchmark. | Stage1 93.0%/92.0%; AIM 94.0%/92.1%: reported gains +1.0/+0.1 points. Reported average SR and percentage-point difference | Reader calculation from the 50 displayed integer task cells gives Hard means of 92.1% for both Stage1 and AIM. The reported Hard gain cannot be reconciled from displayed rows alone; rounding or aggregation details are unspecified. This comparison changes the complete RL stage and does not isolate dense rewards or the attention mask. e-stagee-row-arithmetice-rl |
| Place Mouse Pad Table 1, Easy/Hard RoboTwin task evaluation. | AIM 97%/95%. Task SR | Motus 66%/68%; Stage1 97%/96%. A large baseline advantage coexists with a one-point Hard decrease after RL. The gain over Motus therefore cannot be attributed primarily to post-training. e-task-contraste-protocol |
| Blocks Ranking Size Table 1, Easy/Hard RoboTwin task evaluation. | AIM 47%/43%. Task SR | Motus 75%/63%; X-VLA 67%/74%. The high aggregate result does not imply uniformly successful manipulation or dominance on every task. e-task-contraste-protocol |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. The available stage comparison isolates the complete RL stage, with a much smaller reported gain on Hard. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the two rightmost columns, keeping the Easy and Hard rows separate. Stage1 is the supervised checkpoint; AIM adds action-head RL while freezing the world/value predictor. Then compare external baselines across the same row. AIM reports 94.0% on Easy and 92.1% on Hard, versus LingBot-VA's 92.9% and 91.6%. The bottom row is the paper's combined Average, not another difficulty setting. Treat every number at the precision printed. This table establishes a reported system comparison and a before/after RL comparison; it does not separate the effects of map labels, attention restrictions, dense rewards or additional online experience. e-stagee-aggregatee-protocole-rle-row-arithmetic
What it supports. The reported post-training gains are 1.0 percentage point on Easy and 0.1 on Hard. Stage1 already exceeds the listed external methods in both aggregate settings. The table therefore places most of the reported advantage before RL, but does not identify which supervised architectural or data choice produces it.
Where the evidence stops. No isolated mechanism ablation or uncertainty estimate accompanies this comparison. Table 1's displayed Hard rows do not reproduce the Stage1 aggregate. Also, Table 2 implies a 5.1-point Hard gain over Motus, whereas the accompanying prose says 5.0.
7. Analysis & limitations
7.1 What the evidence leaves open
Figure 2 draws future-frame-to-action and action-to-map arrows, conflicting with the future-RGB → map → action dependence in Equation (10). Its 'PPO Post-Training' label also conflicts with its GRPO caption and Section 4.3. The equations/text specify the intended mechanism; the diagram is not a reliable wiring specification. e-figure-conflicte-maske-rl
The study is simulation-only. No isolated map-removal, mask, reward-component or frozen-branch ablation is reported. Figure 3 contains execution strips without predicted-map overlays or target markers, so it cannot establish the authors' claimed map localization or exclusion of shortcut correlations. e-datae-protocole-qualitative
Evaluation trial counts, seeds, confidence intervals, train/test split details and exact Easy/Hard configurations are missing. These omissions matter especially for the reported 0.1-point Hard post-training improvement. e-protocole-stagee-row-arithmetic
7.2 Questions for discussion
- Would disabling direct RGB access still help when map supervision, parameter count and training budget are controlled? (e-mask; e-protocol)
- Does the frozen map reward improve true task success, or can an action policy exploit erroneous high-value regions? (e-rl; e-stage)
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the Wan2.2 initialization, calibrated three-view simulator data, contact extraction, packed-map encoding and action representation. The paper leaves the Gaussian-width rule, stability threshold, channel semantics and annotation rules beyond pick/place insufficiently specified. e-tokenizatione-data
Training budgets, hardware, optimizer settings, action-head width, loss/reward weights, flow sampling schedule, GRPO group construction and policy-likelihood computation are not supplied. Clarify these alongside the diagram/text conflicts before treating an implementation as an exact reproduction. e-architecturee-losse-maske-rle-protocole-figure-conflict
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test whether the attention restriction contributes beyond map supervision
Reader-proposed experiment, not performed: use the same 30K trajectory split, map labels, pretrained initialization, model size, denoising schedule and supervised budget. Train paired variants with Equation (10)'s action mask versus an otherwise identical mask additionally allowing future-RGB access. First verify the actual attention connectivity, resolving Figure 2 against the equations. Evaluate paired simulation seeds on all tasks and separately on Place Mouse Pad, Scan Object and Blocks Ranking Size, reporting success counts and intervals. If unrestricted access matches or improves success reliably, the claim that forbidding direct RGB is beneficial would be weakened even if map supervision remains useful. e-datae-maske-figure-conflicte-protocole-task-contrast
Check 2: Disentangle map reward from extra online training and verify the Hard gain
Reader-proposed experiment, not performed: start from one Stage1 checkpoint and freeze identical world/value predictors. Compare sparse-only RL, the stated dense-plus-sparse reward, and a control using spatially shuffled maps with the same reward scale, under equal interaction budgets and paired seeds. Keep reward weights and aggregation rules explicit. Measure true task success separately from dense return, including failure cases where the projected target hits a high-value region without completing the task. Preserve raw per-task outcomes and reconstruct both difficulty averages, checking the Table 1/2 discrepancy. A dense-return increase without a reliable success gain, or equal performance with shuffled maps, would undermine the proposed spatial-teacher explanation. e-rle-stagee-row-arithmetice-datae-protocol
8.3 Reading coverage
Visual audit: All eleven original PDF pages were rendered and visually inspected, including the title/byline/version, all method and reward equations, dataset and evaluation text, complete tables, three figures and references. Every final crop was separately viewed; table headers, all task rows, process arrows and Easy/Hard labels are retained. Figure 2's frame/action/map arrows and PPO label were cross-checked against its caption and page 6, with conflicts explicitly disclosed. No appendix appears in this PDF; separate supplements, external code and continuous execution videos were not supplied or inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title, authors, affiliations and Abstract (p. 1)
- 1 Introduction (pp. 1–2)
- 2 Related Work (pp. 2–3)
- 3 Overview (p. 3)
- 4 Method and 4.1 Model Architecture (pp. 4–5)
- 4.2 Intent-Causal Self-Attention (p. 6)
- 4.3 Self-Distillation RL Post-Training (p. 6)
- 5 Dataset and Value-Map Annotation (p. 8)
- 6 Experiments (pp. 8–9); complete Tables 1–2 (pp. 7–8)
- 7 Conclusion (p. 9)
- References [1]–[27] (pp. 10–11)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction's figure omission was addressed by visually inspecting all eleven PDF pages and all five final original crops.
- The inspected source is arXiv:2604.11135v1 [cs.RO], 13 April 2026, marked 'Preprint. Under review.' Its title and all six authors match the catalog; no revision or edition difference was observed. No other version was supplied or compared.
- No appendix is present in the supplied eleven-page PDF. No separate supplement was supplied.
- Code and external resources were not inspected; no training, robot execution or benchmark reproduction was performed.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title/byline, affiliations and arXiv margin stamp
Exact catalog title; authors Liaoyuan Fan, Zetian Xu, Chen Cao, Wenyao Zhang, Mingqi Yuan and Jiayu Chen. Affiliations name INFIFORCE Intelligent Technology Co., Ltd.; The University of Hong Kong; Shanghai Jiao Tong University. Stamp: arXiv:2604.11135v1 [cs.RO], 13 Apr 2026.
Go to primary source ↓e-motivationPDF pp. 1–2, Abstract and Section 1
The authors motivate explicit spatial interaction intent as an interface between dense future appearance and control.
Go to primary source ↓e-overviewPDF p. 3, Figure 1 and Section 3, Equation (1), Architecture Overview
History-conditioned horizon chunks contain future RGB, aligned three-channel maps and actions. The action factor conditions on future maps. Figure 1 adds Future Value and a Self-Distillation arrow to Action; its input diagram labels State and Prop. Tokenizer.
Go to primary source ↓e-tokenizationPDF pp. 4–5, Section 4.1 Tokenization, Equations (2)–(4)
Head/wrist views use a common T-pose layout. RGB/maps share the Wan VAE. The text mentions black-map initialization, an action MLP, T5 features and a history prefix of observations/actions/language.
Go to primary source ↓e-architecturePDF p. 5, Section 4.1 Architecture, Equations (5)–(8), final inference paragraph
Gaussian-noise future streams, learned value-noise token, joint RGB/map flow matching and continuous action decoding. Branches share attention but retain specialized projections and feed-forward layers. Language conditions video; rollout supports cached history and new real observations.
Go to primary source ↓e-lossPDF p. 5, Section 4.1, Equation (9) and explanation
The objective sums RGB flow-matching, weighted map flow-matching and weighted inverse-dynamics action losses; coefficient values and detailed training configuration are absent.
Go to primary source ↓e-maskPDF p. 6, Section 4.2, Equations (10)–(11) and following paragraph
Explicit visibility sets route future RGB into maps and future maps into actions, forbidding direct future-RGB attention by the action stream. Current observations and past actions remain visible to actions.
Go to primary source ↓e-rlPDF p. 6, Section 4.3, Equations (12)–(13)
Only the action head is updated; video/value prediction is frozen. Dense reward samples the predicted map at a camera-projected action target and combines with sparse task feedback. The text names GRPO and provides a clipped probability-ratio objective.
Go to primary source ↓e-dataPDF p. 8, Section 5, all three paragraphs
30K synchronized RoboTwin simulation trajectories; contact projection and adaptive Gaussian smoothing for pick maps; stability detection and support contact for place maps. Numerical annotation thresholds, channel encoding and other task-specific annotation rules are not supplied.
Go to primary source ↓e-protocolPDF pp. 8–9, Section 6, Experimental Setup through Main Results and Analysis; Tables 1–2 on pp. 7–8
50 Easy/Hard simulation tasks, success-rate metric, stated common 30K training data and success criteria; Stage1 is the supervised pre-RL control. No trial counts, uncertainty, detailed split, isolated map/mask/reward controls or hardware/training budget are reported.
Go to primary source ↓e-aggregatePDF p. 8, Table 2, Easy/Hard/Average rows; Section 6 Main Results; p. 7 Table 1 Average row
Table 2 reports AIM 94.0/92.1/93.1, LingBot-VA 92.9/91.6/92.2, Fast-WAM 91.9/91.8/91.8, Motus 88.7/87.0/87.8 and π0.5 82.7/76.8/79.8 percent. Prose claims +5.0 over Motus on Hard, whereas displayed Table 2 entries differ by 5.1 points; Table 1 Motus Hard is 87.02%.
Go to primary source ↓e-stagePDF p. 8, Table 2, Stage1 and AIM columns; Section 6 Baselines and Metrics
Stage1 has 93.0/92.0/92.5 percent Easy/Hard/Average; AIM has 94.0/92.1/93.1. Section 6 identifies initialization from Stage1 and frozen video/value branches.
Go to primary source ↓e-row-arithmeticPDF p. 7, Table 1, all 50 Stage1 Hard and AIM Hard cells and Average row
Reader arithmetic on displayed integer percentages: both Hard columns sum to 4605, yielding an equal-task mean of 92.1%. The printed aggregate is Stage1 92.0% versus AIM 92.1%; raw precision or alternative weighting is not explained.
Go to primary source ↓e-task-contrastPDF p. 7, Table 1, Place Mouse Pad, Blocks Ranking Size, Scan Object, Turn Switch, Hanging Mug and Press Stapler rows
Place Mouse Pad: AIM 97/95, Stage1 97/96, Motus 66/68. Blocks Ranking Size: AIM 47/43, Motus 75/63, X-VLA 67/74. Scan Object: AIM 100/98 versus Motus 67/66. Turn Switch: AIM 100/98 versus Motus 84/78. Hanging Mug: AIM 43/42. Press Stapler: AIM 96/94. All are Easy/Hard percentages.
Go to primary source ↓e-figure-conflictPDF p. 4, Figure 2 graphic and caption; p. 6, Equations (10), (12)–(13)
The graphic includes a Future Frame Generation → Action Head arrow, an Action Head → Spatial Value Map arrow and a PPO Post-Training box. Its caption and Section 4.3 instead say GRPO; Section 4.2 formally excludes direct future RGB from action visibility.
Go to primary source ↓e-qualitativePDF p. 9, Figure 3 and caption; Section 6 continuation and Section 7
Execution strips show place mouse pad, press stapler, scan object, turn switch and open laptop, with Easy left and Hard right. No map overlays, projected-target markers, numerical map-quality diagnostic or predicted-versus-observed comparison is shown. The adjoining prose attributes gains to meaningful map localization and argues against shortcuts.
Go to primary source ↓8.5 Primary sources
AIM: Intent-Aware Unified world action Modeling with Spatial Value Maps ↗
PDF · 5,477 extracted words
Source fingerprint
5fe76983f3e41065f018930c8bae57d7422cdedfc63c565934b706e944aea040