acceptodds
Under review as a conference paper at ICLR 2027

SPARK: Self-Correcting Embodied Memory via Physical Re-Observation

Abstract

Embodied retrieval-augmented generation (RAG) systems build semantic knowledge bases (KBs) from robot observations and answer language queries over them. Most retain and extend the first episode’s KB, even though perception errors can introduce incorrect labels, merged instances, and missing regions from the outset. We present SPARK, a corrective multi-episode framework for embodied RAG. After the first episode, a semantic critic flags implausible and under-covered regions without ground truth, a targeted planner generates revisit viewpoints, and new observations are incorporated through a selective merge policy that can add, remove, or replace information. We evaluate SPARK on four outdoor AirSim maps under two viewpoint degradations using a perception-built KB, a visibility-matched development-time oracle, and a fixed EQA set scored before and after correction. One corrective episode improves downstream question answering on every map under both degradations while storing 19% fewer objects than append-only updating, with gains up to 2.6× on a single map and condition. Extending correction through Episode 4 shows that the resulting KB can exceed the EQA performance of a single undegraded collection pass. Ablations show contributions from the critic, targeted revisits, and selective merge. SPARK shows that physical re-observation can correct persistent robotic memory by identifying weak regions, collecting new evidence, and selectively updating the KB.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied AI, retrieval-augmented generation, semantic memory, knowledge base correction, active perception.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ii4kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents SPARK, a multi-episode framework for correcting persistent embodied memory through physical re-observation. The framework consists of three main components: a semantic critic that identifies unreliable regions in an existing knowledge base, a targeted revisit planner that acquires additional observations, and a selective merge strategy that updates stored information by adding, removing, or replacing object entries. Unlike query-driven memory refinement, SPARK performs memory correction independently of downstream questions.

Critical strengths
  1. The paper addresses an important problem in embodied memory systems: errors and incomplete observations in previously constructed knowledge bases can persist and affect downstream reasoning. Its focus on query-independent memory correction provides a meaningful perspective beyond conventional query-driven exploration and memory retrieval.

  2. By acquiring new observations rather than relying exclusively on previously collected evidence, SPARK can recover information that was missed during initial exploration. The reported improvements in EQA demonstrate the practical potential of physical re-observation for improving persistent embodied memory.

Critical weaknesses
  1. Although SPARK considers a relatively distinct setting of query-independent memory correction, closely related approaches have already investigated uncertainty-guided physical revisitation and reliable persistent memory updating. For example, SCOUT explores semantic uncertainty-driven revisitation for scene-graph refinement, while PBD-AG investigates evidence-based memory maintenance and active inspection. Compared with these approaches, SPARK primarily relies on heuristic diagnostic thresholds, fixed-offset viewpoint generation, and rule-based memory merging. The paper does not sufficiently establish how these mechanisms provide more effective error diagnosis, observation planning, or memory correction than existing alternatives.

  2. Although the paper demonstrates EQA improvements from additional observations, the effectiveness of its core self-correction mechanisms remains insufficiently validated. First, Table 2 shows that selective merging achieves a slightly lower EQA gain than append-only updating (+0.080 vs. +0.082). While it reduces memory size, its advantage in accurately correcting erroneous entries remains unclear. Second, Table 3 reports only critic recall, without meaningful precision evaluation, making it difficult to assess whether the critic can reliably distinguish erroneous memory from correctly reconstructed regions.

  3. The manuscript bears obvious traces of AI-generated writing and appears to contain copied content (e.g., the overuse of em dashes, a typical feature of AI-generated text). Please thoroughly review and revise the manuscript.

Questions for authors
  1. What are the key technical advantages of SPARK over existing uncertainty-guided revisitation and evidence-based memory updating approaches? Can the authors provide direct or capability-matched comparisons to substantiate these advantages?

  2. Given that selective merging does not outperform append-only updating in EQA, how can the authors demonstrate that it more reliably corrects erroneous memory? Can object-level correction accuracy, false deletion rates, and memory consistency be evaluated?

1 comment
AuthorsOpenReview ↗

SCOUT and PBD-AG, merge vs. append, critic precision, writing

Thank you for pointing us to SCOUT and PBD-AG and for the thoughtful questions.

To reiterate the main point of the paper, SPARK adapts the corrective RAG principle to embodied RAG. Prior work such as Mind Palace, ReMEmbR, and Embodied-RAG does not treat the KB as something that can be wrong about what is in the scene. Instead of proposing one more method, SPARK is a whole correction loop, with diagnosis, revisits, merging, and its evaluation, built around an API-based perception system and an LLM/VLM-queryable KB, with the focus on the text that EQA retrieves.

  1. ii4k pointed to SCOUT and PBD-AG.

We will cite both (SCOUT became public in June and PBD-AG in August, shortly before the deadline, while SPARK began long before), but they target different goals and are therefore not directly comparable, although they are complementary and welcome. For example, SCOUT can be categorized as active perception, building online a 3D scene graph of label probabilities and CLIP embeddings, and PBD-AG targets change tracking, following objects that move, appear, or disappear indoors. In contrast, SPARK corrects a text KB that already exists, so that later questions are answered from corrected content, and revisiting is only how it gets the evidence. More specifically, SCOUT relies on a still error-prone detector and its own label confidence, which needs open models like CLIP and produces no captions, while SPARK works with API-based models that produce text and diagnoses the stored KB itself. Main distinction being that neither stores an LLM/VLM-queryable KB from an earlier episode, nor measures whether EQA improves. SPARK's components can be improved, for instance with a modified SCOUT traversal strategy, but a better first pass still cannot rule out errors, and SPARK is about the errors that do get stored. However, we already name better viewpoint selection and navigation as limitations and online correction as future work (Sec. 7).

  1. ii4k noted that the selective merge gains slightly less than append and asked whether it corrects memory, so we also ran deduplicated append offline on the same observations (mean over four maps: append +0.302 with 387 objects per map, selective +0.300 with 262, deduplicated append +0.299 with 265).

At first, these look interchangeable, since the selective merge already performs deduplication, please refer to Algorithm 3 in the Appendix. Getting the same EQA as append is not a weak result, because NLP memory research shows that keeping the full history is the strongest baseline (Mem0, Chhikara et al., 2025, Li et al., 2024). However, a properly queryable RAG needs a healthy KB, and one would not want to keep appending corrupted information to it. From our data append and deduplication keep all 86 objects the critic flagged as implausible across both conditions. The selective merge, on the other hand, removes 46, which is every single one a revisit had in view but could not visually confirm, such as three hallucinated "security cameras", a "sandbag", and two "electrical junction boxes". EQA does not show this, because questions are created automatically only from the oracle view, which could not produce these hallucinations. This removal also depends on the critic, since the merge only removes objects the critic flagged, so without the critic all 86 would stay in the KB, as with append.

  1. ii4k asked about critic precision.

Instead of "undefined", it is rather that precision is 100% and hardly informative, because no region is adequately recovered, even in the undegraded reference memory (where the critic flags 63% of regions), so false positives cannot occur (Appendix C). Regarding the thresholds, like the merge rules, they are simple on purpose and fixed across maps, so that gains and failures can be traced.

  1. ii4k raised concerns about AI-generated writing and copied content.

As our AI use statement says, we used AI tools to help with language editing, while the writing and its manual editing are ours, and the em dashes were our own stylistic choice. The manuscript does not contain copied content, and we would be grateful if the reviewer could point us to any specific passage of concern so that we can check it.

Between submission and now we kept improving the work, and we uploaded a revised PDF. The main change is that the captioning prompt earlier relied on a coarse vocabulary of 22 object types, a common way to match the evaluation vocabulary, and now labels are fully free text, grouped per map. Other changes are (1) the selective merge no longer discards object types seen once in a region, which had removed correct objects, and (2) object placement now uses AirSim's perspective depth instead of the planar formula. The main EQA gain rises from +0.112 to +0.216, as expected after these fixes, and not every change favoured us, since the depth fix lowered the low-lateral gain.

Reviewer_u8ubOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SPARK, a self-correcting framework for embodied scene memory. The method employs a VLM-driven Semantic Critic to identify regions in the knowledge base that contain potential errors or suffer from insufficient observation; it then uses a heuristic planner to generate viewpoints for reobservation and updates the existing knowledge base via Selective Merge. The authors evaluated the method across four outdoor AirSim scenes, reporting improvements in Embodied Question Answering performance following multiple rounds of memory updates.

Critical strengths
  1. Existing embodied memory systems typically focus on the continuous collection and storage of new information, often overlooking the correction of errors within existing memories. Addressing this issue, this paper proposes a method enabling robots to autonomously identify unreliable memory segments and rectify them through re-observation—without relying on ground-truth annotations—thereby offering a viable approach to enhancing long-term memory reliability.
  2. This paper integrates error detection, active re-observation, and memory updating into a comprehensive error-correction workflow. Experimental evaluations were conducted across various AirSim environments and under conditions of degraded viewpoints; furthermore, ablation studies and multi-round error-correction experiments were performed to analyze the roles of individual modules and the method's capacity for iterative improvement. The results demonstrate that the proposed method improves knowledge base quality and enhances performance in downstream scenebased question-answering tasks, thereby validating the framework's effectiveness.
Critical weaknesses
  1. Limited Algorithmic Novelty and Over-Simplified "Embodied" Setup: While applying knowledge bases and RAG to robotics is a promising direction, the proposed pipeline primarily integrates existing off-the-shelf foundation models (e.g., SAM2 and GPT-5-mini) with heuristic filtering, offering relatively limited core algorithmic innovation. More importantly, the experimental setup significantly abstracts away key challenges of embodied AI. Specifically, the agent only records observations at four manually selected waypoints per map, with no sensing during transit. To bypass obstacle avoidance and path planning, inter-waypoint navigation is handled by ascending to a 20m transit altitude, and on Downtown West, continuous flight control is replaced with discrete "pose-stepping" due to tree geometry (Section 5.1, Appendix D.2/D.4). In addition, the study assumes exact ground-truth localization and does not include real-world hardware validation. Because continuous perception, SLAM, and obstacle-aware navigation are excluded from the loop, the current formulation operates much closer to static multi-view image retrieval than to a realistic embodied robotic system.

  2. Marginal Gains from the Selective Merge Policy and Scalability Challenges Across Episodes: A central motivation of the proposed framework is to use a "Selective Merge" policy to filter out erroneous nodes and improve upon standard "append-only" updates. However, the ablation results in Table 2 indicate that Selective merge (ΔEQA = +0.080) yields nearly the same downstream improvement as—and even slightly trails—the simple Append only baseline (ΔEQA = +0.082). This suggests that the observed EQA gains are primarily attributable to acquiring additional viewpoints rather than the selective filtering mechanism itself. Furthermore, the longitudinal study (Section 6.5) highlights a notable scalability limitation: without robust cross-episode object deduplication and identity association, the knowledge base experiences substantial growth over repeated updates. For instance, in the "Old Shipyard" scene (which contains 152 oracle objects), the memory accumulates 1,732 object nodes by Episode 4 (over an 11× increase), accompanied by a slight decline in EQA scores in later episodes on maps such as Hangar and Old Shipyard (Figure 3). This unchecked growth in stored nodes makes it difficult to fully substantiate the claim of a compact, self-correcting persistent memory over extended deployments.

Questions for authors
  1. Hard-Coded Altitude Reset in the Revisit Planner (Algorithm 2, Line 19): Under the low-lateral degradation condition, Episode 1 deliberately lowers the camera height to induce view degradation. However, in Algorithm 2 (Line 19), the revisit planner explicitly sets the Episode 2 altitude back to the "nominal capture height for that waypoint (not the degraded height flown in Episode 1)." Does this imply that the planner has prior knowledge of the undegraded nominal height and manually undoes the vertical degradation by rule? How much of the EQA gain under low-lateral remains if Episode 2 is restricted to the same degraded altitude as Episode 1 with only horizontal offsets?

  2. Multi-Episode Behavior of Merge Policies and Performance Regression (Table 2 & Section 6.5): Given that Append only (ΔEQA = +0.082) slightly outperforms Selective merge (ΔEQA = +0.080) in the singletransition ablation (Table 2), how do Append only and Selective merge compare across Episodes 1 to 4 in Section 6.5 in terms of both EQA scores and total object counts? Furthermore, what causes the noticeable EQA score decline in later episodes on Hangar and Old Shipyard (Figure 3)—is it due to the selective merge erroneously dropping valid nodes, or due to retrieval interference from the massive accumulation of duplicate nodes (e.g., 1,732 nodes on Old Shipyard)?

  3. Waypoint-Bound EQA Queries vs. Global Scene Evaluation (Section 5.3): In the current evaluation protocol, every EQA question is explicitly tied to a single source waypoint using stored provenance (e.g., "How many barrel objects were visible from waypoint 1?"). Since this setup primarily tests localized, per-waypoint observation recovery rather than globally integrated 3D scene memory, how would the system perform on scenewide queries that require aggregating and deduplicating information across multiple distinct waypoints?

2 comments
AuthorsOpenReview ↗

setup, merge vs. append, revisit height, longitudinal results

Thank you for the detailed look at the setup and the concrete questions.

To reiterate the main point of the paper, SPARK adapts the corrective RAG principle to embodied RAG. Prior work such as Mind Palace, ReMEmbR, and Embodied-RAG does not treat the KB as something that can be wrong about what is in the scene. Instead of proposing one more method, SPARK is a whole correction loop, with diagnosis, revisits, merging, and its evaluation, built around an API-based perception system and an LLM/VLM-queryable KB, with the focus on the text that EQA retrieves.

  1. u8ub questioned the novelty and the simplified embodied setup.

The task itself is new, and we use off-the-shelf models on purpose, as the point is the correction loop with a reasonable but error-prone perception model. We also aim to track carefully what correction does and focus on what affects the KB most, the perception components (SAM2, VLM captions, noisy depth). Therefore, the simplified parts such as known pose, a static world, fixed waypoints, and no sensing in transit are deliberate, because with SLAM drift or continuous sensing a gain could simply come from more frames. Regarding falsifiability, despite the controlled parts, many things could have gone wrong, for instance, the critic is never told how Episode 1 was degraded and never sees the oracle, the planner can miss a region, and the merge can delete valid objects. Despite all this, every map improves under both conditions.

  1. u8ub noted that the selective merge does not outperform append, so we also ran deduplicated append offline on the same observations (mean over four maps: append +0.302 with 387 objects per map, selective +0.300 with 262, deduplicated append +0.299 with 265).

At first, these look interchangeable, since the selective merge already performs deduplication, please refer to Algorithm 3 in the Appendix. Getting the same EQA as append is not a weak result, because NLP memory research shows that keeping the full history is the strongest baseline (Mem0, Chhikara et al., 2025, Li et al., 2024). However, a properly queryable RAG needs a healthy KB, and one would not want to keep appending corrupted information to it. From our data append and deduplication keep all 86 objects the critic flagged as implausible across both conditions. The selective merge, on the other hand, removes 46, which is every single one a revisit had in view but could not visually confirm, such as three hallucinated "security cameras", a "sandbag", and two "electrical junction boxes". EQA does not show this, because questions are created automatically only from the oracle view, which could not produce these hallucinations.

  1. u8ub asked whether the planner undoes the degradation by restoring the nominal height.

It does restore it, by design, since the planner flies the height and pitch the mission planned for each waypoint. The robot is never told that Episode 1 was degraded or how, it only has its own memory and images, so restoring the height is not knowledge of the error but the standard way to observe the region again. Restoring the height only adds new views, and with append or deduplication new views only add objects, so all 86 implausible objects would stay, while the critic is what decides where to look and what to remove. Both the guided and untargeted arms restore the height, so their comparison isolates the critic's choice of regions (+0.132 vs. +0.118), and they are close for two reasons:

  • After degradation the critic flags 62 to 81% of clusters, so the untargeted arm also spends most of its budget on flagged ones.
  • 112 of the 158 answers are identical between the arms, so the difference comes from the few regions the guided arm reaches first.
  1. u8ub asked about the later decline on Hangar and Old Shipyard.

In the revision Hangar improves at every episode (0.438 to 0.553), and Old Shipyard rises from 0.200 to 0.558 by Episode 3. It is true that growth is still an issue, but it is a stated limitation (Sec. 6.5), and the longitudinal study uses the selective merge only.

  1. u8ub asked about scene-wide queries.

We form questions around one waypoint's region on purpose, so that each loss stays tied to its own region instead of being averaged over all regions.

Between submission and now we kept improving the work, and we uploaded a revised PDF. The main change is that the captioning prompt earlier relied on a coarse vocabulary of 22 object types, a common way to match the evaluation vocabulary, and now labels are fully free text, grouped per map. Other changes are (1) the selective merge no longer discards object types seen once in a region, which had removed correct objects, and (2) object placement now uses AirSim's perspective depth instead of the planar formula. The main EQA gain rises from +0.112 to +0.216, as expected after these fixes, and not every change favoured us, since the depth fix lowered the low-lateral gain.

Thank you to the authors for the explanations and clarifications. I find the justification regarding the work's novelty quite convincing; with these clarifications, I now understand how this approach differs from Mind Palace, ReMEmbR, and Embodied-RAG. However, given the nature of the experimental results—specifically the large overall sample size and the potential influence of various "tricks" employed—I cannot be certain that the observed performance gains stem solely from the proposed method itself. Consequently, I can only raise my score to "Weak rejection."

Reviewer_PCquOpenReview ↗
Rating
2
Confidence
2(+1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces SPARK. SPARK repairs embodied memory through semantic diagnosis, targeted physical revisitation, and selective updates that add, remove, or replace stored objects. Experiments on four outdoor AirSim maps under two viewpoint degradations use 151 fixed EQA questions. One corrective episode improves the average score from 0.366 to 0.478, while selective merging retains fewer objects than append-only updating with a similar EQA gain.

Critical strengths
  • The paper addresses a practically relevant problem: errors introduced during perception can persist in robotic memory and affect subsequent tasks.
  • The diagnosis, revisitation, and update stages form a coherent workflow, supported by component ablations and evaluation on the same questions before and after updating.
  • Critic-guided revisitation outperforms untargeted revisitation under the same observation budget on all three maps where degradation reduces initial performance.
Critical weaknesses
  • W1. The contribution of error correction is not sufficiently isolated from additional observation. In Table 2, selective merging and append-only updating achieve EQA gains of +0.080 and +0.082, respectively. Retaining fewer objects demonstrates potential storage efficiency, but does not establish that incorrect records are repaired. Object-level measurements are needed to distinguish successful correction from information recovery and unintended deletion of valid content.

  • W2. The evaluation does not establish reliable diagnostic discrimination. Section 6.3 reports no negative examples, preventing assessment of false alarms on adequately reconstructed regions. In addition, the oracle is described as pose-matched in Section 4 but uses undegraded baseline poses according to Section 5.2. This distinction mixes perception errors within observed views with information unavailable from degraded views, complicating interpretation of what the critic detects.

  • W3. Long-term reliability remains insufficiently demonstrated. Figure 3 shows Hangar declining from approximately 0.575 at Episode 2 to 0.500 at Episode 4, below its undegraded reference of 0.525. Old Shipyard also accumulates 1,732 stored objects versus 152 in the oracle reference. Although acknowledged by the authors, these limitations directly affect persistent memory maintenance. Repeated judge scoring does not establish robustness across independent executions of the complete pipeline.

Questions for authors
  • Q1: Under identical revisit observations, can the authors compare selective merging with deduplicated append-only updating and report label-correction, false-object removal, and valid-object deletion rates?
  • Q2: Can the authors evaluate the critic on memories containing both reliable and unreliable regions, separating errors relative to the actually observed views from missing information relative to the fixed baseline reference?
  • Q3: Can the authors provide independently repeated, budget-matched longitudinal comparisons, including memory growth and uncertainty estimates? To support reproducibility, please also reconcile the planar-depth conversion in Appendix B.3 with the perspective-depth input specified in Appendix D.1, which AirSim distinguishes explicitly.
1 comment
AuthorsOpenReview ↗

deduplication and removals, critic precision, depth fix and stability

Thank you for the careful reading and thoughtful requests.

To reiterate the main point of the paper, SPARK adapts the corrective RAG principle to embodied RAG. Prior work such as Mind Palace, ReMEmbR, and Embodied-RAG does not treat the KB as something that can be wrong about what is in the scene. Instead of proposing one more method, SPARK is a whole correction loop, with diagnosis, revisits, merging, and its evaluation, built around an API-based perception system and an LLM/VLM-queryable KB, with the focus on the text that EQA retrieves.

  1. PCqu mentioned whether the selective merge corrects or only adds views and how deduplication compares, hence we ran the requested (W1, Q1) deduplicated append offline on the same observations (mean over four maps):
  • Append: +0.302, 387 objects per map
  • Selective merge: +0.300, 262
  • Deduplicated append: +0.299, 265

At first, these look interchangeable, since the selective merge already performs deduplication, please refer to Algorithm 3 in the Appendix. Getting the same EQA as append is not a weak result, because NLP memory research shows that keeping the full history is the strongest baseline (Mem0, Chhikara et al., 2025, Li et al., 2024). However, a properly queryable RAG needs a healthy KB, and by definition one would not want to keep appending corrupted information to it. For example, if a drone scanned a dock and a harbour operator asked how many fuel drums are left on the pier, they would not want a count that hallucinations made too high or too low, and this is not something append or deduplication deal with. We agree that we should have measured this better, so we added an error management analysis to the revised paper (Sec. 6.2). From our data append and deduplication keep all 86 objects the critic flagged as implausible across both conditions. The selective merge, on the other hand, removes 46, which is every single one a revisit had in view but could not visually confirm, such as three hallucinated "security cameras", a "sandbag", and two "electrical junction boxes". Another 37 were never in a revisit's view and stay until one is.

So why does EQA not show it? Because questions are created automatically only from the oracle view, which could not produce the hallucinated security cameras, sandbags, and electrical junction boxes. This removal also depends on the critic, since the merge only removes objects the critic flagged, so without the critic all 86 would stay in the KB, as with append.

  1. PCqu asked about the oracle poses being known and the critic's precision being undefined (W2, Q2).

Both are by design, because the oracle is occlusion-aware, field-of-view-restricted simulator ground truth from the undegraded poses, showing what practically perfect perception would capture and therefore which questions can be asked, and our real-perception run from the same poses is used as a check that collection is reasonable at all. You would be right to ask "Why have degradations then, instead of simply relying on the still imperfect undegraded view?" They are a controlled error, without which we would rely on marginal changes from real perception that already starts from a good (intended) view. Instead of "undefined", it is rather that precision is 100% and hardly informative, because no region is adequately recovered, even in the undegraded reference memory (where the critic flags 63% of regions), so false positives cannot occur (Appendix C).

  1. PCqu questioned the stability of long-term behaviour, a possible depth formula mismatch, and variance (W3, Q3).

In the revision Hangar improves at every episode (0.438 to 0.553), and three of four maps exceed their undegraded reference by Episode 3, with Old Shipyard within 0.004 of it. And even with 13 to 46% fewer objects per map than append, growth is still a stated limitation (Sec. 6.5). On variance, six repeated Episode 1 collections stayed within 0.05, and judge passes differ by 0.004 (Appendix C). Regarding the mistake of applying the planar-depth formula to AirSim's perspective depth, the reviewer is definitely right, and we also caught the mismatch earlier, corrected it, and re-ran everything for the revision.

Between submission and now we kept improving the work, and we uploaded a revised PDF. The main change is that the captioning prompt earlier relied on a coarse vocabulary of 22 object types, a common way to match the evaluation vocabulary, and now labels are fully free text, grouped per map. Other changes are (1) the selective merge no longer discards object types seen once in a region, which had removed correct objects, and (2) object placement now uses AirSim's perspective depth instead of the planar formula. The main EQA gain rises from +0.112 to +0.216, as expected after these fixes, and not every change favoured us, since the depth fix lowered the low-lateral gain.

Comments on the paper

1 comment
Area_Chair_aDiBOpenReview ↗

Reminder: Please Review Authors’ Rebuttal and Provide Feedback

Dear Reviewers,

Thank you for your reviews. If you have not already done so, please read the authors' rebuttal and share your feedback in the discussion thread at your earliest convenience. Your timely responses would be greatly appreciated.

Best,
AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.