Paper Reading Notes (Deep): PACA: Perspective-Aware Cross-Attention Representation for Zero-Shot Scene Rearrangement

1. Basic information

• Abbreviation: PACA
• Authors: Shutong Jin, Ruiyu Wang, Kuangyi Chen, Florian T. Pokorny
• Affiliations: KTH Royal Institute of Technology, Graz University of Technology
• Venue: WACV 2025
• Published paper: CVF Open Access
• Reading mode: Deep Dive

2. Key assessment

PACA turns Stable Diffusion generation into an object-level perception interface. Goal generation exposes cross-attention for prompt words; real-image DDIM inversion / reconstruction exposes comparable attention. Middle denoising maps provide regions and late maps provide finer features. Their combination supports real–generated object matching, relative transformations, and robot execution.

The evidence has four levels:

1. Goal quality: 43 participants and 559 ratings place PACA’s three scene averages near those of two human arrangers;
2. Prompt design: among 2,000 images per scene, 6,000 total, category-list prompts beat exact-count and top-down descriptions on IS / CLIP score;
3. Object matching: on 450 real–generated pairs, joint cross-attention averages 86.72%, versus 51.35% for DALL·E-Bot’s Mask R-CNN + CLIP pipeline;
4. Robot execution: three scenes, two settings, and 60 real trials yield 68.73% / 65.56% average success for 3-DoF / 6-DoF.

For an object-pose collection, four boundaries must remain explicit:

• The main text parameterizes transformations as $(x,y,z,\theta)$, without full roll / pitch / yaw;
• 6-DoF mainly describes oblique observations, perspective-controlled goals, and the robot system; it does not imply standard $SE(3)$ object-pose output;
• Execution success measures relative layout agreement, allowing slight absolute-position errors and lacking pose GT;
• The clearest quantitative gain is object matching, rather than rotation / translation accuracy.

PACA belongs among extended systems that integrate transformations into generative robotics, rather than BOP / ADD(-S) 6D pose methods.

2.1 Evidence matrix

Experiment	Sample	Metric	Supports	Does not establish
User study	43 people, 559 ratings	Satisfaction, 1–5	Human-preferred goals in small everyday scenes	Geometric executability, correct counts, or all-user preference generalization
Prompt comparison	6,000 generated images	IS, CLIP score	Less distortion and better text agreement for category prompts in this SD setup	More controllable precise spatial relationships
Matching	450 image pairs	Object matching accuracy	Fewer matching errors with a joint representation	More accurate numerical object transformations
Robot execution	60 trials	Relative-layout success	A real system completing about two-thirds of tasks	Industrial-precision SE(3) pose, grasping, or placement

3. Problem setting: generate goals and recover transforms

3.1 Scene rearrangement

Let the initial real scene be $X$ and the human-preferred goal be $X^*$. An embodied agent transforms scene objects to approach $X^*$:

$$X\xrightarrow{\{T_i\}}X^*.$$

Conventional approaches learn desirable layouts from training data. Dining, Fruit, and Office have different relationships and user preferences; fixed simulated patterns may also produce a reality gap.

PACA uses a web-scale generator for image goals and converts them into object actions. This replaces learning the goal distribution with three interface challenges:

1. Do generated objects follow the prompt?
2. Can generated and real objects be matched?
3. Can a 2D goal supply metric transformations for the robot?

3.2 Why multistage errors accumulate

Let $G,S,E,M$ denote the generator, segmenter, encoder, and matcher, respectively. The conventional path is:

$$x^{goal}=G(P),\quad \{m_i\}=S(x^{goal}),\quad f_i=E(x^{goal}\odot m_i),\quad \pi=M(\{f_i\},\{f_j^{real}\}).$$

If $G$ distorts, duplicates, or adds objects, $S$ produces repeated / misplaced crops, and $E$ compresses wrong crops into features; lost image information cannot be recovered. Errors alter downstream input distributions rather than merely adding independently.

PACA directly uses internal prompt-word / spatial-token cross-attention as segmentation and feature information, reducing model interfaces.

3.3 What zero-shot means

Training-free / zero-shot means no new network training or robot demonstrations for the three rearrangement scenes. The system still depends on:

• Web-scale Stable Diffusion image–text training;
• ControlNet conditional generation;
• Marigold monocular-depth priors;
• DDIM inversion;
• ICP geometry optimization;
• RealSense RGB-D and robot calibration.

Zero-shot concerns task adaptation; learned priors, calibration, and a manually constructed pipeline remain necessary.

4. What PACA means by 6-DoF

4.1 Three DoF contexts

1. 3-DoF top-down setting: near-overhead goal and real views, mainly planar translation / rotation;
2. 6-DoF setting: oblique real-camera perspectives, requiring view-compatible goals and depth;
3. 6-DoF robot arm: UFactory Lite 6 has six joint degrees of freedom.

None automatically implies object pose with three-axis $SO(3)$ rotation plus 3D translation.

4.2 Object transformations in the main text

Section 4.5 explicitly gives:

$$T=(x,y,z,\theta).$$

This is 3D translation and one rotation scalar. With top-down suction and tabletop placement, $\theta$ is planar orientation / yaw. No three independent rotational degrees in $R\in SO(3)$ or mechanism for flipping arbitrary sideways objects into full target poses is reported.

Figure 2 labels execution $(R,T)$ and the conclusion mentions 6-DoF operations, but these broad labels do not override the specific method parameterization. This site uses the most explicit mathematical definition for task scope.

4.3 Evaluation is not a pose benchmark

Success means final relative positions agree with the generated goal, allowing slight absolute-position variations. The paper provides no:

• translation error (cm / mm);
• rotation error (degree);
• ADD / ADD-S;
• VSD / MSSD / MSPD;
• 6D ground-truth object pose;
• per-object transformation residual.

Thus Table 4’s 65.56% is task success, not 6D pose accuracy.

5. Overall PACA architecture

Four stages:

1. Input a prompt and real RGB-D;
2. Generate a perspective-compatible goal and object representations;
3. Invert / reconstruct real RGB, extracting matching real representations;
4. Execute grasping and placement through depth, ICP, and robot geometry.

This combines foundation models, attention maps, geometric alignment, and scripted grasping rather than an end-to-end policy.

6. Thread A: prompt-to-goal generation

6.1 3-DoF top-down generation

Prompt $P$ lists object categories; random seed $s$ reproduces the goal:

$$x^{goal}=\operatorname{StableDiffusion}(P,s,\beta_{cfg}),$$

Classifier-free guidance is fixed at:

$$\beta_{cfg}=7.5.$$

The authors favor category lists over exact one-to-one counts and articles:

fork, knife, spoon, plate, table

Instead of:

A fork, a knife, a plate, and a spoon, top-down

Experiments show higher IS / CLIP scores for the former under this Stable Diffusion distribution.

6.2 From top-down to perspective-controlled goals

A key image-goal challenge is viewpoint mismatch: an oblique wrist camera versus a generated top-down view. Inspired by perspective lines in drawing, PACA extracts Hough lines from real RGB:

$$C_h=\operatorname{HoughTransform}(x^{real}).$$

These line conditions control generation:

$$x^{goal}=\operatorname{ControlNet}(C_h,P,s,\beta_{cfg}).$$

Table edges, wall corners, and furniture contours supply perspective cues, making generated viewpoints closer to the real scene.

6.3 This does not explicitly recover camera pose

Although the text describes aligning goal / real extrinsic matrices with Hough lines, the implementation conditions image generation without reporting:

• camera intrinsics;
• Numerical real / goal extrinsics;
• reprojection error;
• line-to-vanishing-point calibration;
• camera pose ground truth.

A more precise interpretation is perspective appearance conditioning rather than geometric calibration. Views may still mismatch without stable lines, when lines come from objects rather than the environment, or when ControlNet ignores conditions.

7. Thread B: real images in diffusion representation space

7.1 Inputs and latents

Real RGB $x^{real}$ is resized to 512×512 and encoded by Stable Diffusion $E$ into clean latent:

$$z^{real}_0=E(x^{real}).$$

Ordinary generation starts from noise. To read real-image cross-attention on the same denoising timeline, first find a noise latent reconstructing the image.

7.2 DDIM inversion

Deterministic DDIM inversion maps $z^{real}_0$ from $t=1$ to $t=T$, yielding noise $z^{real}_T$. Standard denoising then reconstructs the real image with CFG 0, reducing prompt-induced content changes.

This supplies:

1. A hidden process using the same U-Net / timesteps as goal generation;
2. Cross-attention between prompt words and real image regions.

Inversion builds an attention-readable reconstruction trajectory rather than generating a different semantic scene from real RGB.

7.3 Possible errors

• Inversion / reconstruction is not strictly lossless;
• Stable Diffusion latents may poorly represent tiny industrial objects or nonnatural scenes;
• Real objects absent from the prompt lack word-specific representations;
• Synonyms, compound words, and tokenization affect attention;
• CFG=0 removes text-guidance strength without guaranteeing pixel-identical reconstruction.

8. Joint cross-attention representation

8.1 Cross-attention definition

The paper uses:

$$M=\operatorname{Softmax}\left(\frac{QK^T}{\sqrt{d}}\right),$$

where:

$$Q=W_Q\phi_z(z_t),\qquad K=W_Kp,\qquad V=W_Vp,$$

$p=\psi(P)$ is the prompt embedding; $M(w,t)$ is the spatial attention map for word $w$ at timestep $t$. Maps are upsampled / represented at 512×512, with activations in 0–1.

8.2 Information levels during lossy denoising

The authors compare denoising to progressive decoding: large shapes / regions emerge at high noise, followed by texture and local features. Figure 4’s plate / fork example shows:

• Middle-step attention covering whole silhouettes;
• Late attention focusing on discriminative fork heads and plate patterns.

They define:

• Conceptual embedding (region): $t\approx T/2$;
• Perceptual embedding (feature): $t\approx 1$.

Total denoising steps:

$$T=50.$$

8.3 Constructing the representation

Threshold the two attention maps separately:

$$R(w)=\operatorname{Step}(M(w,T/2),\tau_1)+ \operatorname{Step}(M(w,1),\tau_2),$$

where:

$$\tau_1=0.3,\qquad \tau_2=0.9.$$

The lower region threshold retains full object regions; the higher feature threshold retains strongest late discriminative activations while reducing matching computation.

8.4 Segmentation and feature encoding as functional roles

The two terms of $R(w)$ play segmentation-like and feature-like roles. This is not:

• A segmentation mask trained with pixel GT;
• A conventional CLIP embedding with high-dimensional appearance vectors;
• A dense descriptor trained with explicit correspondence supervision;
• A panoptic representation with instance IDs.

It is thresholded word-conditioned attention support. Generation and representation are naturally synchronized, but depend on tokens, thresholds, and attention interpretability.

9. Object matching and transformation estimation

9.1 Matching objective

For each word $w$, PACA solves between real / goal representations:

$$T_{opt}(w)=\arg\min_T \left\|R^{goal}(w)-T(R^{real}(w))\right\|^2.$$

ICP estimates relative alignment, initialized from real representations. These mainly encode 2D regions / feature support; depth is still needed for robot actions.

9.2 Goal depth

Generated RGB lacks metric depth. Marigold estimates $D_{est}$, then real RGB-D depth $D_{real}$ fits scale / shift parameters:

$$\mathcal{L}(\vartheta)= \left\|D_{est}(\vartheta)-D_{real}\right\|_2^2.$$

$\vartheta$ denotes depth scaling / shifting, distinct from object yaw $\theta$. Aligning relative monocular goal depth to real depth approximates placement z, but generated sizes, perspectives, and hallucinated shapes still affect it.

9.3 Final object transformations

The main text parameterizes:

$$T=(x,y,z,\theta).$$

This provides 3D placement and planar orientation. Sideways cutlery, tilted cups, or objects needing flipping lack a full execution mechanism in this representation and top-down suction strategy.

10. Robot execution

10.1 Platform

• UFactory Lite 6;
• wrist-mounted Intel RealSense D405 RGB-D camera;
• vacuum gripper;
• Before grasping, the camera is above the table;
• Representation geometric centers are grasp points;
• Grasping is top-down.

The robot has 6 joint degrees of freedom, while its tabletop grasp / place strategy remains highly structured.

10.2 Scenes

Scene	Objects	Main relationships
Dining	Knives, forks, spoons, plate	Cutlery around a plate
Fruit	Apples, oranges, plate	Fruit on / near a plate
Office	Keyboards, mice, mug	Relative office-supply arrangements

These reuse DALL·E-Bot’s everyday tabletop setting for baseline comparison. Small object counts and regular workspaces do not represent open-world tidying.

10.3 Handling generated extra objects

Stable Diffusion may add unrequested objects. Because PACA builds representations only for prompt words, extra objects lack word maps and do not enter execution. This reduces matching to extras but does not fix missing objects or wrong same-category counts: a category token does not naturally give two apples stable individual IDs.

11. Experiment I: goal satisfaction

11.1 Protocol

• 43 subjects;
• 559 ratings;
• Ratings from 1 (very bad) to 5 (very good);
• Human 1, Human 2, PACA, Parallel, and SG-Bot;
• Dining / Fruit / Office scenes.

11.2 Results

Method	Dining	Fruit	Office
Human 1	3.41	4.63	3.55
Human 2	4.34	1.89	3.67
PACA	3.85	4.42	3.49
Parallel	3.05	2.26	1.95
SG-Bot	1.68	—	—

PACA approaches Human 1 on Fruit and lies between / near the two humans on Dining / Office. Human 2 scores only 1.89 on Fruit, highlighting subjective ideal layouts rather than one uniquely correct GT.

SG-Bot lacks applicable trained models / simulated objects for Fruit / Office. Missing entries cannot serve as numerical evidence of PACA outperforming it.

11.3 Evaluation boundaries

Satisfaction evaluates visual layouts without measuring:

• Grasp feasibility;
• Collisions;
• Object stability;
• Realistic physical sizes;
• Robot reachability;
• Consistency with individual user preferences.

12. Experiment II: prompt forms

12.1 Protocol

Generate 2,000 images per scene, 6,000 total. Compare categorical versus specific one-to-one descriptions using:

• Inception Score, used by the authors to reflect distortion / generation quality;
• CLIP Score: image–prompt alignment.

12.2 Complete numerical results

Scene	Prompt	IS	CLIP
Dining	fork, knife, spoon, plate, table	4.08±0.21	31.19
Dining	A fork, a knife, a plate, and a spoon, top-down	3.54±0.24	30.33
Fruit	apple, orange, plate	3.53±0.19	21.54
Fruit	Two apples, and an orange, top-down	2.99±0.15	19.80
Office	keyboard, mouse, mug	3.81±0.20	20.54
Office	A keyboard, a mouse, and a mug, top-down	3.55±0.21	20.32

The authors suggest LAION-5B captions more often describe categories than exact inventories, making categorical prompts more familiar to Stable Diffusion.

12.3 A clear trade-off

Category prompts improve overall image quality while reducing:

• exact object count;
• one-to-one identity;
• precise spatial relation;
• user personalization.

The conclusion lists object-count / spatial-relationship control as future work. Higher IS / CLIP scores should not obscure weaker control.

13. Experiment III: joint-representation matching accuracy

13.1 Protocol

• 150 real–generated pairs per scene;
• 450 pairs total;
• Naturally lit real images with texture variation;
• Stable Diffusion goals generated from the same categories;
• Count successfully matched objects relative to expected matches per pair, aggregating across 150 pairs.

DALL·E-Bot uses Mask R-CNN segmentation plus CLIP encoding; PACA uses joint diffusion cross-attention representations.

Method	Dining	Fruit	Office	Average
DALL·E-Bot	42.83%	42.57%	68.66%	51.35%
PACA	83.75%	88.63%	87.78%	86.72%

Absolute gains:

• Dining: +40.92 percentage points;
• Fruit: +46.06 points;
• Office: +19.12 points;
• Average: +35.37 points.

The relative average improvement is about 68.9%, but percentage-point differences are clearer here.

13.2 Why is the Office baseline higher?

Table 3 provides no rigorous attribution. Larger keyboard / mouse / mug appearance differences and stronger Mask R-CNN / CLIP familiarity are plausible explanations, not validated conclusions.

13.3 Persuasive evidence with limits on extrapolation

The experiment directly targets multistage error accumulation and improves all three scenes, but does not ablate:

• Changing segmentation alone;
• Changing CLIP features alone;
• conceptual map only;
• perceptual map only;
• threshold sensitivity;
• timestep sensitivity;
• SD version / layer aggregation.

It validates the joint design as a whole without isolating every internal contribution.

14. Experiment IV: real-robot execution

14.1 Protocol

• 3 scenes;
• 3-DoF and 6-DoF settings;
• 10 random object initializations per scene per setting;
• $3\times2\times10=60$ trials total;
• No object-pose ground truth;
• Success: relative positions align with the goal, allowing slight absolute-position variation.

Setting	Dining	Fruit	Office	Average
3-DoF	61.93%	64.61%	79.67%	68.73%
6-DoF	56.67%	66.67%	73.33%	65.56%

14.2 Interpreting differences

• The 3.17-percentage-point average drop is modest;
• Dining / Office decline in 6-DoF, while Fruit improves;
• The paper mainly attributes 6-DoF failures to generated-goal depth errors;
• 60 trials is a small sample without confidence intervals;
• Percentages are not simple multiples of 10 trials, suggesting object-level / multiple-criterion averaging. The text does not break this down, so exact success counts must not be inferred.

14.3 Possible failure propagation

Prompt count / relationship errors
          ↓
Generated-object distortion or viewpoint mismatch
          ↓
Cross-attention region / feature displacement
          ↓
Real–goal matching or ICP error
          ↓
Marigold depth scale / shift error
          ↓
Geometric center is not a feasible grasp point / center of mass
          ↓
Suction, movement, or placement failure

Table 4 gives only final success, without stage contributions. Future work should log stage-specific failure labels.

15. Strengths

15.1 One model bridges real and generated domains

Real and goal appearances differ substantially. Instead of requiring an external segmenter / encoder to fit both, PACA obtains word-conditioned attention from the same U-Net during generation and reconstruction, creating a shared interface.

15.2 Strong semantic context

Prompts identify apples, oranges, or plates. Category words can ignore instance appearance differences across real / generated domains more readily than color histograms, local keypoints, or templates.

15.3 Fewer module interfaces

A joint representation does not make the whole system simple: foundation and geometry modules remain. It does remove the key sequence of generic segmentation, patch cropping, and encoding for generated images.

15.4 Real execution

Beyond attention visualizations and matching benchmarks, robot execution exposes actual limits in depth, grasp centers, and goal control.

16. Limitations and unresolved questions

16.1 Incomplete pose definition

The central inclusion boundary: $(x,y,z,\theta)$ is not complete $SE(3)$, with no pose GT / error metric. Do not rank PACA alongside standard 6D methods such as FoundationPose, MegaPose, or GDR-Net.

16.2 Imprecise goal constraints

Category prompts improve quality without reliably controlling:

• Object counts;
• Left / right, front / back, and contact relationships;
• object identity;
• Reachability and collisions;
• Physical stability;
• Personal preferences.

The paper suggests energy-based models for better goal control.

16.3 Perspective control is a weak geometric condition

Hough lines encode only 2D structure, without guaranteeing matching extrinsics, object scales, or principal points. Curved tables, line-free backgrounds, clutter, narrow fields of view, and occluded lines may cause failures.

16.4 Depth is a bottleneck

Marigold gives relative depth; global affine scale / shift fitting assumes it can be aligned with real depth. Different generated shapes / scales leave local metric errors that one alignment cannot correct.

16.5 Weak grasp planning

Geometric centers serve as vacuum grasp points, but:

• They can fall in holes or nonsuctionable regions;
• They are not centers of mass;
• Surface normals / suction seals are unchecked;
• Collision / reachability scores are absent;
• Transparent, flexible, porous, or heavy objects are not handled.

The conclusion also notes missing grasp-critical information such as center of mass.

16.6 Multiple instances

Word attention aggregates by category. Two apples or identical utensils may form multiple connected regions without stable instance indexing / one-to-one assignment. Generated count changes further complicate this.

16.7 Scale and statistics

• 3 scenes;
• 450 image pairs;
• 60 robot trials;
• No diversity of robots / cameras;
• No confidence intervals;
• No stage-specific failure counts;
• No standard pose benchmark.

The results establish feasibility rather than reliability at scale.

16.8 Reproduction materials

No verifiable official code link appears in the main text / CVF page. Training-free reproduction still requires SD / ControlNet layer extraction, DDIM inversion, attention aggregation, thresholds, ICP, depth alignment, camera–robot calibration, and manipulation integration.

17. Reproduction checklist

17.1 Perception / generation

• Fix Stable Diffusion version, scheduler, and 50 denoising steps;
• CFG 7.5 for goal generation;
• CFG 0 for real reconstruction;
• Save random seeds;
• Fix 512×512 real-RGB preprocessing;
• Record U-Net layers, heads, and attention aggregation;
• region timestep $T/2$, feature timestep 1;
• thresholds 0.3 / 0.9;
• Specify tokenization of plurals, compounds, and synonyms;
• Record Hough settings and ControlNet checkpoint.

17.2 Matching / geometry

• Initialize each object from $R^{real}(w)$;
• Specify connected components / instance separation;
• Record ICP source / target points, distance thresholds, and iterations;
• Fix Marigold version;
• Specify pixels used for scale / shift fitting;
• Handle missing depth, reflection, and occlusion;
• Convert image coordinates to camera / robot frames;
• Verify units and coordinate directions.

17.3 Robot

• Calibrate Lite 6 and D405 extrinsics;
• vacuum tool center point;
• top-down approach / retract height;
• workspace safety margin;
• Check that grasp centers have valid masked depth;
• Check placement collisions;
• Plan failure recovery;
• Save prompts, seeds, goals, attention maps, depth, ICP transforms, and execution logs for every trial.

17.4 Better evaluation

Alongside task success, add:

• Matching precision / recall and Hungarian assignment;
• mask IoU;
• camera-view / vanishing-point error;
• Depth AbsRel / RMSE where synthetic or calibrated GT is available;
• planar translation / yaw error;
• Complete $SE(3)$ error if roll / pitch are added;
• Separate grasp, placement, and layout success;
• stage-wise failure taxonomy;
• Confidence intervals over multiple random seeds.

18. Potential improvements

These extensions arise from the limitations and are not completed paper contributions.

18.1 Explicit camera conditioning beyond Hough lines

Condition viewpoints directly on calibrated intrinsics / extrinsics, depth edges, or scene geometry, rather than indirectly through image lines. Camera parameters could condition ControlNet / adapters, evaluated by reprojection consistency.

18.2 Scene graphs before images

Language or energy models could generate verifiable graphs: nodes contain categories / counts; edges encode left-of, inside, near, aligned, and similar relations. Check counts and physical constraints before image generation.

18.3 Instance-aware attention decomposition

Create instance proposals from same-category connected regions, then enforce one-to-one matches with geometry, depth, and assignment, avoiding mixed apples / forks in one token map.

18.4 Depth-aware representations

The paper also suggests integrating depth estimation into representation construction. Align attention with metric real RGB-D clouds for 3D support; use generated multiview / depth uncertainty for goals rather than treating monocular depth as certain.

18.5 Complete SE(3) and grasp-aware goals

To enter the core 6D pose collection, the method would need:

• Outputs $R\in SO(3)$ and $t\in\mathbb{R}^3$;
• Sideways / flipped object handling;
• Symmetry-aware pose metrics;
• Goal grasp surfaces, center of mass, stable support, and collision constraints;
• Multiview / 3D goal representations beyond one image.

19. Relationship to the object-pose research map

PACA can be located as:

Language prompt
    ↓
Stable Diffusion goal generation and object cross-attention
    ↓
Real RGB-D DDIM inversion / reconstruction
    ↓
Real–goal object matching
    ↓
ICP and depth alignment recover (x,y,z,θ)
    ↓
Robot grasp / place and relative-layout evaluation

Overlap with core 6D pose research:

• Object-level correspondence;
• Relative transformation;
• RGB-D and geometric alignment;
• Transformations used in robot actions.

Differences:

• Goals come from generation rather than real references / CAD;
• Main output is a rearranged scene;
• Rotation is only $\theta$;
• No pose GT or standard metric.

Extended inclusion can demonstrate downstream pose value while preserving the integrity of standard pose-paper statistics.

20. Deep-reading assessment

PACA offers a useful foundation-model interface: diffusion cross-attention forms comparable object representations in goal generation and real reconstruction, reducing segmentation–crop–encoder errors. 86.72% versus 51.35% over 450 pairs is its strongest direct evidence.

Robot execution establishes more than attention visualization, but 68.73% / 65.56% task success also reveals bottlenecks in generated counts / relations, monocular depth, grasp centers, and small-sample reliability.

The most precise assessment is:

PACA is a zero-shot image-goal rearrangement pipeline using perspective-conditioned generation and joint cross-attention for object matching and $(x,y,z,\theta)$ transformations. Real-robot relative-layout success is about two-thirds. It is not a complete $SE(3)$ object-pose estimator and does not report standard 6D pose accuracy.

Future work should incorporate camera geometry, instance counts, metric depth, full rotation, grasp affordances, and stage-specific uncertainty into representation and evaluation, rather than broadening 6-DoF terminology.