SG-WAM: Text-Grounded and Spatial-aware Semantic Guidance for World-Action Models
1. Paper overview
In one sentence: Predict what the instruction means for the future scene, then let that foresight guide video and action together. e02e03e10e11e13e15
| At a glance | What to know |
|---|---|
| Research problem | Author claim The authors argue that observation-independent instruction embeddings let video-based manipulation models follow visual salience rather than the requested object or destination. SG-WAM seeks to make predicted dynamics and actions instruction-specific by conditioning generation on future features inferred jointly from the current scene and instruction. e02 |
| Core mechanism | |
| A key reported result | Robustness to perturbations: 81.3% overall; 81.7% under language perturbation. Success rate over perturbed tasks. LIBERO-Plus; seven factors perturbed individually. GE-Act overall: 77.8%, a 3.5 percentage-point gap. OpenVLA-OFT language: 79.5%, a 2.2 percentage-point gap. Reported robustness gains exceed the standard-benchmark margin. This protocol tests individual perturbations rather than their combinations. e11 |
| Reading caution |
Core contributions
Figure 2. A future-feature planner guides two coupled generative experts. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at panel (a), with the current observation, instruction and three groups of query tokens. Their hidden states become forecasts of language-aligned and geometry-aware features. Panel (b) supplies the training targets: frozen teachers see actual future frames, while the planner only sees the present. Follow the blue path into panel (c): semantic guidance enters the video expert; its joint attention with the action expert carries that information into action generation. The text encoder remains a separate conditioning path. The top row depicts supervision, so those future targets should not be mistaken for observations available to the deployed policy. e03e04e05e06
What it supports. The method has two distinct levels of future prediction: compact semantic foresight followed by joint video/action generation. Its action mechanism is not a separate inverse-dynamics model applied to a finished video. The architecture explains why calling the entire system a single undifferentiated predictor would conceal the planner–generator interface.
Where the evidence stops. The figure summarizes connectivity. Section 3.3 is needed to establish that the additional semantic cross-attention is injected into the video expert and reaches the action expert indirectly. It does not specify the real-time replanning schedule.
2. Motivation
2.1 The problem and the proposed response
The authors argue that observation-independent instruction embeddings let video-based manipulation models follow visual salience rather than the requested object or destination. SG-WAM seeks to make predicted dynamics and actions instruction-specific by conditioning generation on future features inferred jointly from the current scene and instruction. e02
2.2 What this reading follows
A robot can imagine a visually plausible movement and still reach for the wrong object. SG-WAM frames this as an instruction-grounding problem: a text embedding prepared independently of the scene may leave the video model free to follow visual salience. The proposed remedy is a planner that looks at both the present observation and the instruction, then forecasts language-aligned and geometry-aware features of future keyframes. This reading follows that information through the generative policy and tests the explanation against the paper's original architecture, benchmark tables, physical-trial chart and component ablations. e02e03e10e11e13e15
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Joint prediction |
| Quadrant | Q3 · Dual-system × Joint prediction |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded Q3 classification is supported by a distinct VLM semantic planner feeding a world-action executor with coupled video and action experts. Equation (1) specifies joint future-observation/action prediction, and action generation uses joint attention rather than an explicit inverse-dynamics decoder. Semantic foresight remains an inference-time input, beyond its auxiliary alignment loss. This assessment rests on the architecture and information flow, not joint training alone. e03e06e07e08
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 Forecast feature maps, not a verbal plan
The word planner can suggest a chain of language subgoals, but SG-WAM's planner produces dense feature maps. Qwen3.5 2B receives an image, an instruction and learnable queries. A shared base group captures information common to both branches; branch-specific groups specialize toward language-aligned or geometric future features. A resampler attends to both image states and query states, so the output can retain spatial detail while anticipating change. The implementation predicts four future keyframes, each represented by 256 spatial tokens. The two targets have different feature dimensions: 1024 for SigLIP2 and 2048 for Depth Anything 3. These are teacher representation spaces, not discrete object labels or raw metric-depth predictions. e04e05e09
5.2 Keep the teacher's future out of the deployed inputs
Future-frame teachers are available during learning because demonstrations contain later observations. The planner is trained to reproduce those features using only the current image and instruction. The target encoders stay frozen, and the targets are detached. Crucially, the video/action system consumes the planner's predictions during co-training, not the teacher's actual future features. At deployment the teachers disappear. This distinction prevents the architecture diagram's supervision arrows from being mistaken for access to ground-truth futures at test time. The three training stages first establish foresight, then teach a video generator to use it, and finally introduce action learning while retaining the alignment objective. e03e05e08
5.3 Locate the world-to-action interface
The two forecast branches are projected into a shared guidance space and fused through a learned scalar gate. The video expert receives this representation through an additional cross-attention branch, while retaining its existing text conditioning. The action expert receives robot state and communicates with the video expert through joint attention. The resulting model predicts nine future frames and a 36-step action chunk. My reading is that the important world–action connection is the shared generative information flow: the system does not first finalize a video and then infer an action through a separate inverse-dynamics stage. This still leaves an operational gap: chunk length alone does not tell us how many actions execute before feedback and replanning. e03e06e07e09e12
5.4 Training and inference
During training
Stage 1 trains the VLM, query embeddings and resamplers against frozen future-frame teachers: SigLIP2 penultimate-layer patch tokens and Depth Anything 3 last-layer features. Spatial supervision matches features rather than raw depth. e05e08e09
Stage 2 jointly trains planner and video expert with L_vid + λ_plan L_plan. Stage 3 adds the action expert and optimizes L_gen + λ_plan L_plan. Generation is conditioned on planner predictions during training, avoiding substitution of teacher features for deployment inputs. e08
During inference
Teacher encoders are discarded. One planner forward pass supplies foresight for generation; the same planner is applied independently to each view. Guidance dropout during training enables classifier-free guidance at inference. e03e04e06
Predicting an action chunk is distinct from executing it. The real-world evaluation establishes physical use, but the supplied method does not specify how many chunk steps execute before observation refresh or replanning. e03e09e12
5.5 Implementation flow
- Forecast future representations
Qwen3.5 2B processes each camera view independently with shared weights. It appends 32 base query tokens and 32 tokens per semantic branch to the observation–instruction sequence. Each branch reads shared plus specific hidden states, reusing tokens across four evenly spaced future keyframes. e04e09
- Recover dense spatial features
A resampler's learnable grid queries attend to both image states and semantic query states. Each output map contains 256 tokens, with feature dimensions 1024 for text-grounded semantics and 2048 for spatial-aware semantics. e04
- Condition the video expert
Projected feature branches are fused with a learned scalar gate. Time and spatial positional encodings locate the guidance. Every video-expert block receives an additional parallel cross-attention branch while retaining its original text cross-attention. e06
- Generate coupled video and actions
An LTX-Video backbone and an action expert jointly predict future video and action chunks. The action expert receives robot state and accesses semantic guidance indirectly through joint attention with the video expert. The described action mechanism is joint generative prediction, without an explicit inverse-dynamics stage. e03e06e07e09
6. Experiments & results
SG-WAM adds a VLM planner that forecasts dense semantic and geometric features of future observations. These features condition a video expert, while an action expert receives their influence through joint attention. The reported advantage is small on standard LIBERO but larger under LIBERO-Plus perturbations. Physical trials support the approach separately from illustrative generated-video examples; the bar chart confirms the reported ordering but supplies no printed exact success rates.
6.1 Read the original evidence
Table 1. Standard LIBERO is near saturation; the average does not tell the whole story. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the four suite columns before the rightmost average. SG-WAM is strongest in Spatial and Goal, but FastWAM leads Object and LingBot-VA leads Long. The Type column marks the paper's VLA/WAM grouping; it does not imply identical training budgets. The protocol uses 500 evaluation trials per suite. Compare the last row with LingBot-VA to assess the strongest average baseline, then with GE-Act and FastWAM, which also appear in the physical evaluation. Bold denotes the best value and underlining the second best in the original table. e10
What it supports. SG-WAM reports 98.7% average success versus 98.5% for LingBot-VA, a 0.2 percentage-point difference. It does not win every suite. The more informative question is whether its semantic interface helps when the scene or instruction shifts, which motivates Table 2.
Where the evidence stops. The table does not give confidence intervals or repeated-training variability. A small difference near the benchmark ceiling should not be interpreted as established statistical superiority or as a universal benefit on every manipulation task.
Table 2. Robustness gains are larger than the standard-benchmark gain, but remain uneven. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each column changes one factor: Layout, Camera, initial robot state, Language, Light, background texture or Noise. Compare within a column, because the interventions pose different challenges. The average aggregates the perturbed tasks; do not recompute it as an unqualified arithmetic mean of the seven displayed columns. SG-WAM's Language value is 81.7%, while its Camera value is 54.8%. These cells are useful for testing the intended mechanism: improved semantic grounding does not automatically solve geometric changes in the observation process. e11
What it supports. The overall result is 81.3%, versus GE-Act's 77.8%, a 3.5-point gap. Language improves by 2.2 points over OpenVLA-OFT's 79.5%. SG-WAM nevertheless trails pi-zero-FAST in Camera and OpenVLA-OFT in background texture. The evidence supports broad average improvement across this suite, not uniformly best robustness.
Where the evidence stops. These are factor-specific perturbations, not a test of every combined shift. Cross-method comparisons also cannot isolate the guidance mechanism as cleanly as a matched ablation within SG-WAM. The camera weakness remains a useful reproduction target.
Figure 4. Physical manipulation is evaluated separately from generated video. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The first four groups cover kiwi placement, ordered bowl stacking, ordered pan filling and bimanual pot lifting. The final two test a previously unseen basket height and changed lighting. Match the blue, green and coral legend to GE-Act, FastWAM and SG-WAM. The paper reports 100 demonstrations per task and 50 evaluation trials per model and task, with 50 trials again in each generalization setting. The chart makes relative ordering visible, while the surrounding task description establishes that these are physical robot experiments. e12e13
What it supports. SG-WAM has the tallest bar in every displayed group, including the height and lighting shifts. This is direct physical-execution evidence within the stated AgileX task suite, complementing the simulation results. It should be read separately from the DROID video examples below.
Where the evidence stops. The bars do not carry exact numerical labels, so this report does not invent exact success percentages from pixel heights. The figure also does not display error bars. Four task families and two shifts leave many forms of deployment variation untested.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Standard simulated manipulation LIBERO Spatial, Object, Goal and Long suites; 500 evaluation trials per suite. | 98.7% Average task success rate | LingBot-VA: 98.5%; FastWAM: 97.6%; GE-Act: 96.5%. Author-reported highest average among listed methods. The 0.2 percentage-point lead over LingBot-VA is small; uncertainty is not reported. e10 |
| Robustness to perturbations LIBERO-Plus; seven factors perturbed individually. | 81.3% overall; 81.7% under language perturbation. Success rate over perturbed tasks | GE-Act overall: 77.8%, a 3.5 percentage-point gap. OpenVLA-OFT language: 79.5%, a 2.2 percentage-point gap. Reported robustness gains exceed the standard-benchmark margin. This protocol tests individual perturbations rather than their combinations. e11 |
| Semantic guidance ablations LIBERO; authors state the same data setting and identical hyperparameters. | Full: 98.7%; without guidance: 97.2%; without text-grounded semantics: 97.7%; without spatial-aware semantics: 98.3%; without shared queries: 98.4%. Average task success rate | Drops from full model: 1.5, 1.0, 0.4 and 0.3 percentage points, respectively. Both branches contribute in this setup, with a larger decrement from removing text-grounded features. These aggregate scores do not directly measure target-selection versus geometric errors. e15e16 |
| Training schedule ablation LIBERO; single-stage versus staged training. | Single-stage: 97.9%; full staged model: 98.7%. Average task success rate | 0.8 percentage-point advantage for staged training. Supports the schedule empirically. The explanation that staging supplies more reliable foresight before action learning is the authors' interpretation, without a reported reliability trajectory. e15e16 |
| Physical manipulation and generalization AgileX Cobot; four tasks with 100 demonstrations per task and 50 test trials per task/model; two additional generalization settings with 50 trials each. | Exact percentages are not printed; the original chart shows the highest success rate in all four standard tasks and both generalization settings. Executed-task success rate | GE-Act and FastWAM; raised basket height and changed lighting evaluated after standard-setting training. This concerns physical execution. Figure 4 was visually checked for model ordering, but its unlabeled bar heights were not converted into exact percentages. e12e13 |
| Instruction following in predicted futures DROID-trained models; identical reference frames with instructions unseen during training. | Authors describe correct bottle destination and carrot selection in two SG-WAM examples. Qualitative correspondence to instructed object/destination | The RGB-based WAM is described as choosing the stove and the wrong item. These are generated-video examples, not executed-robot success measurements or an aggregate instruction-following benchmark. The original example images were visually inspected. e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. Qualitative instruction following makes the intended benefit visible. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. For each instruction, read across the three frames in the top and bottom rows. The top row is the RGB-based WAM and the bottom row is SG-WAM. In the left example, the requested destination is the microwave rather than the stove. In the right example, the instruction specifies the yellow carrot in the sink and a black pot. The original circles and markers highlight the selected object or location. These are generated futures from the same reference frame under instructions described as unseen during training; they are not a second table of physical robot trials. e14
What it supports. The examples illustrate why a scene-conditioned semantic forecast could be more useful than an instruction embedding alone: the intended object and destination constrain which plausible future is relevant. They make the qualitative failure mode concrete and motivate the component removals in Table 3.
Where the evidence stops. Selected generated-video examples do not establish the frequency of correct grounding, causal mediation through the planner, or successful execution of those imagined motions. A held-out test with systematically swapped targets would provide stronger evidence for that mechanism.
Table 3. Removing guidance hurts most; the branch and staging gains are smaller. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the full model's 98.7% as the reference, then subtract each row independently. Removing all semantic guidance costs 1.5 percentage points. Removing text-grounded semantics costs 1.0, removing spatial-aware semantics costs 0.4, replacing shared query tokens costs 0.3, and single-stage training costs 0.8. The source groups these changes by semantic guidance, dual semantics and training stages. Section 4.5 says the variants use the same data setting and identical hyperparameters. The rows remove or replace different components, so their deltas are not separate terms that can be added together. e15e16
What it supports. Within the reported setting, every tested change reduces average LIBERO success. The largest removal supports the guidance interface as a useful addition; the text-grounded branch has a larger observed removal cost than the geometry branch. Staged learning is also helpful in this comparison.
Where the evidence stops. These are single reported rates without uncertainty estimates. The table does not establish that small branch gains are robust across training seeds, or that the same ordering holds under LIBERO-Plus perturbations. It does not separate added capacity from every other training difference.
7. Analysis & limitations
7.1 What the evidence leaves open
The main text does not establish matched training data and compute across all benchmark baselines or supply confidence intervals. The small LIBERO lead therefore does not establish statistically reliable superiority. e09e10e11
Future-image feature matching is an indirect grounding objective: the teachers supply visual features rather than explicit correct-versus-incorrect instruction supervision. Ablations and qualitative examples support usefulness but do not isolate causal instruction dependence. e05e14e15
The attention study is descriptive, and the physical generalization tests cover selected height and lighting changes. These results leave broader compositional instruction changes and combined environmental shifts unresolved. e11e12e14
7.2 Questions for discussion
- Would swapping only the instruction produce appropriately different foresight and executed actions on identical scenes? [e05, e14]
- Do spatial features primarily improve contact precision, target localization or robustness to viewpoint changes? [e05, e15]
- How much of the staged-training gain survives equal total optimization budgets and repeated seeds? [e15, e16]
8. Reproducibility audit
8.1 Requirements and known gaps
The reported setup uses eight H100 GPUs: 20k planner-training steps followed by two 30k-step stages. Key settings are λ_spa=0.004, λ_plan=0.25, λ_act=1.0 and guidance dropout 0.15. The implementation paragraph also specifies learning rates and warmup schedules. e09
Reproduction requires the named pretrained backbones, future-frame teacher extraction and aligned video/action demonstrations. The supplied text leaves exact checkpoint revisions, batch sizes, per-stage data composition, action/state parameterization, denoising steps, guidance scale and execution cadence unspecified. e03e05e08e09e12
A useful first experiment would repeat full versus no-guidance training with matched optimization budgets and multiple seeds, then test paired instructions on identical scenes while separately scoring target selection and manipulation completion. e14e15e16
The primary paper text and selected original PDF pages were reviewed. The title page links a project page, but no project code was inspected and no experiments were reproduced. e01
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: Swap the instruction while holding the scene fixed
Construct held-out scene pairs with identical observations and instructions that change only the target object or destination. Compare the full model, no semantic guidance and the text-only branch with matched data and training budgets. Record target selection in predicted futures and executed actions separately. The semantic-grounding explanation would be stronger if gains persist across paired instruction swaps, rather than appearing only in selected video examples or average task success. e02e14e15
Check 2: Test the geometry branch under controlled camera shifts
Repeat the full-versus-no-spatial-branch comparison under LIBERO-Plus camera and layout perturbations, reporting per-condition results and variation across repeated training runs. Keep the text branch, action horizon and evaluation episodes fixed. This would test whether the small standard-LIBERO gain from geometry grows under a relevant shift, and whether it addresses the camera weakness visible in Table 2. No such additional experiment was run for this report. e11e15e16
8.3 Reading coverage
Visual audit: Visually inspected the original PDF pages and all six final crops: architecture Figure 2, Tables 1–2, physical-trial Figure 4, generated-video Figure 5 and ablation Table 3. Headers, legends, axes and numerical cells were preserved. Exact physical success rates were not digitized because Figure 4 has no printed bar values. The method text on pages 4–5 and implementation on page 7 informed the tutorial; external videos, code and robot experiments were not inspected or reproduced.
PDF pages inspected for this edition: 3, 6, 7, 8, 9. Appendix coverage: not established.
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, abstract and version identifier
- 1 Introduction
- 2 Related Work
- 3 Methodology
- 3.1 Overview
- 3.2 Semantic Foresight Learning
- 3.3 Semantic-Guided World-Action Model
- 3.4 Three-Stage Training Paradigm
- 4 Experiments
- 4.1 Experimental Setup
- 4.2 Comparisons with State-of-the-Art Methods (Q1 & Q2)
- 4.3 Multi-task Experiments in the Real World (Q1 & Q2)
- 4.4 Instruction-Following Analysis (Q3)
- 4.5 Ablation Study (Q4)
- 5 Conclusion
- References
- Supplied captions for Figures 1–6
- Independent visual audit of PDF pages 3-9: architecture, equations, Tables 1-3 and Figures 2-6
Outside the original text pass
- Figure 1 was reviewed through its caption and surrounding text only; Figures 2-6 were subsequently inspected in rendered PDF pages 3, 7 and 8.
- Figure 4 was visually inspected and confirms the reported bar ordering, but its bars have no printed numeric labels. Exact real-world percentages are therefore not transcribed.
- The original layout of Tables 1-3 was subsequently visually inspected, resolving the extraction cell-merging issue for the reported results.
- No appendix appears in the supplied 13-page text; separate supplemental material availability is unverified.
- Identity matches the catalog title, identifier and author list. The observed source is arXiv:2608.08839v1, dated 9 August 2026.
- 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 visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, author list, project link and arXiv version line
Title, identifier and authors agree with the catalog; the page identifies arXiv:2608.08839v1, 9 August 2026, and links a project page.
Go to primary source ↓e02PDF pp. 1–2, Section 1 Introduction
The authors attribute instruction misalignment to observation-independent text embeddings and propose scene-conditioned semantic and geometric foresight.
Go to primary source ↓e03PDF p. 3, Figure 2 (visually inspected); p. 4, Section 3.1, Eq. (1)
A VLM planner supplies future feature maps to coupled video/action experts. The joint distribution includes observation, instruction, state and foresight; teachers are removed at inference.
Go to primary source ↓e04PDF p. 4, Section 3.2, Eq. (2)
Shared and branch-specific query states feed resamplers attending to image states. Views share planner weights. Maps have 256 tokens with dimensions 1024 and 2048.
Go to primary source ↓e05PDF pp. 4–5, Section 3.2, teacher descriptions and Eq. (3)
Frozen SigLIP2 and Depth Anything 3 encode actual future frames. Detached targets supervise text-grounded MSE and spatial smooth-L1 feature alignment.
Go to primary source ↓e06PDF p. 5, Section 3.3, Eq. (4) and following paragraph
A scalar-gated sum fuses projected features. Additional video cross-attention retains text conditioning; action attention receives guidance indirectly. Guidance dropout enables classifier-free guidance.
Go to primary source ↓e07PDF p. 5, Section 3.3, interpolation definition and Eq. (5)
Both experts regress flow velocities ε−y from interpolated clean targets and noise; video and weighted action losses form L_gen.
Go to primary source ↓e08PDF p. 5, Section 3.4, Stages 1–3 and Eqs. (6)–(7)
Training progresses from planner alignment to planner/video co-training and then action co-training. Teachers remain frozen, and generation consumes predicted foresight.
Go to primary source ↓e09PDF p. 7, Section 4.1, Implementation Details
Specifies Qwen3.5 2B, LTX-Video, teacher layers, four keyframes, 32 base and 32 queries per branch, nine frames, 36 actions, eight H100s, training schedules and loss/dropout settings.
Go to primary source ↓e10PDF p. 6, Table 1 all suite and Avg columns, visually inspected; Section 4.1 LIBERO protocol; p. 8, Section 4.2
Evaluation uses 500 trials per suite. Reported averages are SG-WAM 98.7%, LingBot-VA 98.5%, FastWAM 97.6% and GE-Act 96.5%. The cited table cells were also checked in the original PDF layout. SG-WAM suite rates are 99.4, 99.4, 98.2 and 97.8. FastWAM leads Object at 100.0 and LingBot-VA leads Long at 98.5.
Go to primary source ↓e11PDF p. 6, Table 2 all factor and Avg columns, visually inspected; pp. 6–7, benchmark description; p. 8, Section 4.2
Seven factors are perturbed individually. Overall rates are SG-WAM 81.3% and GE-Act 77.8%; language rates are SG-WAM 81.7% and OpenVLA-OFT 79.5%. The cited table cells were also checked in the original PDF layout. Camera: SG-WAM 54.8 versus pi-zero-FAST 65.1. Background: SG-WAM 88.9 versus OpenVLA-OFT 93.3. The caption defines Avg as success over all perturbed tasks.
Go to primary source ↓e12PDF p. 8, Section 4.3, Setup and Tasks
AgileX dual-arm experiments use front and wrist RGB cameras, four tasks, 100 demonstrations per task and 50 trials per task/model, plus height and lighting tests.
Go to primary source ↓e13PDF pp. 8-9, Section 4.3, Results; p. 7, Figure 4 (visually inspected)
The original chart confirms that SG-WAM bars are tallest in each of the four standard tasks and both generalization settings against GE-Act and FastWAM. The bars have no printed numerical labels; exact values were not digitized.
Go to primary source ↓e14PDF p. 9, Section 4.4, both qualitative studies; p. 8, Figures 5-6 (visually inspected)
Authors describe two DROID generated-video examples under unseen instructions and more concentrated action-to-video attention on instruction-relevant objects. The original generated-video panels and attention heatmaps were visually inspected; they remain qualitative illustrations.
Go to primary source ↓e15PDF p. 9, Table 3, all variant rows
LIBERO success rates: full 98.7%; no guidance 97.2%; no spatial semantics 98.3%; no text-grounded semantics 97.7%; no shared queries 98.4%; single-stage 97.9%. The cited table cells were also checked in the original PDF layout.
Go to primary source ↓e16PDF p. 9, Section 4.5, protocol and component interpretations
Ablations use the same data setting and identical hyperparameters. Authors associate branches with object selection and geometry, and staging with reliable guidance before action learning.
Go to primary source ↓8.5 Primary sources
SG-WAM: Text-Grounded and Spatial-aware Semantic Guidance for World-Action Models ↗
PDF · 6,354 extracted words
Source fingerprint
bca2bbd1aacb95708a148ec9ff036237e35da87dc3e273d4f66cc8bffed2b0cb