Pith. sign in

REVIEW 5 major objections 6 minor 27 references

This paper claims that a service robot can estimate accident-prone regions in household scenes by propagating data-derived risk scores through a scene graph, reaching 75% agreement with human judgments of whether a scene is risky.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A semantic graph framework propagates risk scores derived from a national accident database across spatial object relations, reporting 75% binary risk detection accuracy on 20 human-annotated NYU V2 home images.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Genuinely new idea, but the 75% accuracy claim is not supported by the evaluation; fixable with straightforward changes. the 5 major comments →

arxiv 2508.19788 v1 pith:VAQC3PAT submitted 2025-08-27 cs.RO cs.CV

Context-Aware Risk Estimation in Home Environments: A Probabilistic Framework for Service Robots

Sena Ishii , Akash Chikhalikar , Ankit A. Ravankar , Jose Victorio Salazar Luces , Yasuhisa Hirata This is my paper
classification cs.RO cs.CV
keywords risk estimationservice robotsscene graphrisk propagationhousehold accidentsRGB-D perceptionprobabilistic frameworkhome safety
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a home service robot can infer which regions of a room are accident-prone from a single RGB-D image, including hazards no single object label would reveal. It builds a scene graph whose nodes are detected objects and whose edges carry two quantities: how often two objects co-occur in real accident reports (e.g., towel near heater leading to fire), and how close they are in 3D space. Risk scores propagate only from higher-risk to lower-risk objects, so a benign towel near an oven can become flagged as a fire hazard. The authors test this on 20 NYU Depth V2 scenes with human annotations and report 75% binary risk-detection accuracy, with the clearest gains after propagation in cut-related scenes. If correct, the framework gives robots a lightweight, transparent way to anticipate household dangers without handcrafted rules or large language models.

Core claim

The central claim is that contextual risk in a domestic scene can be estimated by combining object-level accident statistics with asymmetric spatial-semantic propagation on a scene graph. Each detected object starts with a risk score for three accident types—cut, fire, and trip/fall—derived from a national database of real accident reports using Laplace-smoothed frequency ratios. Risk then flows from a higher-risk object to a lower-risk neighbor with weight equal to the pair's accident co-occurrence correlation times a spatial proximity factor, and only when the source is riskier than the target. The paper reports that this procedure classifies whether a scene contains risk with 75% accuracy

What carries the argument

The engine is an asymmetric risk-propagation algorithm over an object-centric scene graph. Node i's risk ri is updated from neighbor j only when rj > ri, with weight wij = φaccrel[i,j] · (1 − φdistance[i,j]) · (rj − ri), where φaccrel is the Laplace-smoothed co-occurrence accident ratio from the database (pair risk correlation) and φdistance is the Euclidean distance between detected bounding-box centroids (closer neighbors dominate). The asymmetry—risk flows only from dangerous to benign objects—is what keeps the system from inflating risk everywhere and gives the propagation its interpretable direction.

Load-bearing premise

The result stands on the assumption that the Japanese accident database's pairwise co-occurrence statistics are meaningful for the exact object pairs that Detic detects in NYU scenes, and that Euclidean distance between bounding-box centroids captures the physical proximity that makes those statistics relevant.

What would settle it

Recompute the propagation with the accident-correlation term replaced by random edge weights (or with spatial distance set to a constant) and measure binary risk accuracy on the same 20 images: if accuracy stays near 75%, the semantic database term is not doing the claimed work. Alternatively, restrict evaluation to images where detected object pairs have zero co-occurrence records in the database; if accuracy collapses there, the mapping from detections to database keys is the load-bearing step rather than the propagation itself.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A robot using this framework can flag objects as hazardous that a detector alone would call safe, because risk arrives from neighboring objects (e.g., towel near oven).
  • The same pipeline can be extended to other accident types without building new models, since only the database query and risk category change.
  • Because every propagated score traces back to a visible object pair and an accident statistic, the robot can explain its alert (e.g., 'towel near heater').
  • The lightweight graph update is designed to run onboard in real time, making proactive alerts feasible on mobile home robots.
  • The 75% accuracy on binary risk presence suggests that human-level gist of 'is this scene risky?' can be recovered from object layout alone, at least on benchmark home images.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper never specifies the functional form of φdistance or how Detic class names map to database query keys; a natural test is to vary the distance kernel and mapping and measure sensitivity of the 75% figure.
  • The largest gain appears in cut scenes while fire and fall improve less, suggesting object-state cues (pose, heat, occlusion) carry information the current static-object representation cannot capture.
  • The database's counterintuitive entries (e.g., blankets and cut accidents) hint that raw co-occurrence counts conflate causation with correlation; a causality-filtered version might change which pairs propagate risk.
  • A sharper spatial metric than centroid distance—such as contact area, support relations, or reachability by a human hand—could improve fall and fire predictions without changing the propagation rule.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes a framework for estimating accident-prone regions in home scenes for service robots. Objects detected by Detic are assigned initial risk scores for three accident types (cut, fire, trip/fall) using Laplace-smoothed statistics from a Japanese governmental accident database. Risk is then propagated over an object-centric scene graph: a lower-risk node receives influence from higher-risk neighbors, weighted by an accident-correlation term phi_accrel and a spatial-proximity term phi_distance, with updates clipped and renormalized. Heatmaps are generated from smoothed object masks and compared with human annotations. The method is evaluated on 20 NYU Depth V2 images annotated by 14 participants via a GUI; the authors report 75% binary risk/no-risk accuracy and state that propagation improves risk-type classification, especially for cut-related scenes.

Significance. The conceptual direction is timely: interpretable, lightweight risk perception from RGB-D images is relevant for domestic service robots, and the use of real-world accident statistics is a genuine strength. The asymmetric propagation idea and the collection of human risk annotations are also worthwhile contributions. However, the central empirical claims are not supported by the reported evaluation. Section IV.A describes a label-conditioned protocol in which the ground-truth risk type is revealed before the binary decision, so the 75% accuracy does not measure the claimed ability to detect the presence of risk without knowing its type. The decision threshold is never reported, the sample size is 20, and no uncertainty or baselines are given. In addition, phi_distance is used in Eq. (6) but never defined, and the mapping from Detic classes to accident-database keys is unspecified. As submitted, the method is not fully reproducible and the validation does not establish the paper's main claims.

major comments (5)
  1. [Section IV.A] The binary risk-detection evaluation is invalid for the stated task. The text says 'predictions were made using only the heatmap corresponding to the risk type annotated for each image.' The research question posed in Section IV is whether the system can tell whether a scene contains any accident risk without being told the risk type. Conditioning on the annotated risk type reveals the label and bypasses the hardest part of the problem: deciding which heatmap, if any, indicates risk. The 75% accuracy therefore measures only whether a known-type heatmap exceeds a threshold. Re-evaluate using the maximum activation over all three heatmaps (or an equivalent any-risk decision rule), and report the full confusion matrix and the threshold.
  2. [Section IV.A / Fig. 6] The 'predefined threshold' is never specified anywhere in the manuscript. With only 20 images and majority-vote labels, the result is not reproducible and may reflect threshold fitting on the same data. No majority-class baseline, confidence interval, cross-validation, or leave-one-out estimate is reported. At minimum, provide the exact threshold, the complete 2x2 counts underlying Fig. 6, bootstrap uncertainty, and comparison with trivial baselines such as always-risk or majority-class.
  3. [Section III.B / III.C, Eq. (6), Algorithm 1] The paper uses phi_distance[i,j] in the propagation weight but never gives its functional form. The statement that it is 'computed using the Euclidean distances between object centroids' is not a definition; the kernel, normalization, and parameter values are absent. Consequently, the propagation algorithm is underspecified and the experiments cannot be reproduced. Also, Algorithm 1 re-normalizes all risk scores by min-max scaling every iteration (line 17) while the convergence criterion uses max_diff computed before renormalization; the effect of this rescaling on risk ordering and on the stopping condition needs clarification.
  4. [Section III.A / V] The object-to-database mapping is unspecified. The method queries the accident database with object names, but the paper does not explain how Detic class names (e.g., 'cabinet', 'blanket', 'box') are matched to database keys, nor what risk score is assigned when a detected object has no accident records. The Discussion itself notes counterintuitive entries (cut accidents involving blankets). If matching is ad hoc or many objects have zero counts, the initial risk scores and phi_accrel values are arbitrary and the propagation output is not grounded. Specify the matching procedure, the fallback policy for missing objects, and a sensitivity analysis over the smoothing parameter and missing-data handling.
  5. [Section IV.B] The claim that 'a clear improvement was observed after propagation, especially in cut-related scenes' is not supported by quantitative results. The confusion matrices in Fig. 7 are not accompanied by per-class accuracy, precision/recall, or any statistical comparison. The paper's own Discussion says fire and trip/fall performance was less consistent. Additionally, the centroid inverse-distance metric is reported only as 1/d with no numerical values, and 1/d is undefined when d=0. Without numbers and significance testing, the central claim that propagation improves risk classification cannot be evaluated.
minor comments (6)
  1. [Section III.A, Eq. (3)] Notation is inconsistent: Eq. (3) uses N(knife, cut) and N(knife) where Eq. (2) uses count(o,a) and total(o), and N is also used for the number of accident types. The example also omits the Laplace smoothing term introduced in Eq. (2).
  2. [Section IV] Typo: 'tirp/fall' should be 'trip/fall'.
  3. [Table I] The caption does not define the checkmark/cross symbols used in the table.
  4. [Section IV.B, Eq. (7)] The inverse-distance metric 1/d is unbounded as d approaches 0. Add a small epsilon and report the distribution of d values, not just a single aggregate.
  5. [Section IV] No data or code availability statement is given. For reproducibility of the user study, specify the 20 image identifiers, participant recruitment details, and any exclusion criteria.
  6. [General] The paper uses 'semantic alignment' to describe what is actually a centroid-distance metric. Clarify the terminology to distinguish semantic type alignment from spatial alignment.

Circularity Check

1 steps flagged

Binary accuracy is computed with the ground-truth risk type revealed (Sec. IV.A), so the headline 75% does not measure the claimed risk-presence detection.

specific steps
  1. self definitional [Section IV.A, 'Evaluation of Risk Detection Accuracy' (paragraphs 2–3)]
    "An image was predicted as risky if any of the algorithm-generated heatmaps—corresponding to cut, fire, or trip/fall categories—exhibited activation above a predefined threshold. If no such activations were present, the image was labeled as non-risky. ... It is important to note that for this evaluation, predictions were made using only the heatmap corresponding to the risk type annotated for each image. This risk-type-specific evaluation avoids artificial performance inflation..."

    The stated task is to decide whether an image contains any accident risk. This protocol first says 'any' heatmap above threshold, then overrides it: for each image, predictions use only the heatmap corresponding to the ground-truth risk type. For risky images, the system is therefore handed the correct risk type (cut/fire/trip-fall) before thresholding and never has to select the risk type or determine whether a risk exists at all. The 75% accuracy is a threshold check on the label-selected channel, not an evaluation of risk-presence detection. The rule is undefined for no-risk images, which have no annotated type, so the negative class is not actually tested. The threshold is never reported. The prediction is conditioned on the target label by construction, so the headline result does not

full rationale

The core derivation is not circular: object-level risk scores come from an external accident database, propagation uses accident-pair statistics and spatial distance, and human annotations are independent of the algorithm. Self-citations appear mainly in related work and are not load-bearing for the central derivation. However, the central empirical claim—75% binary risk detection accuracy—rests on an evaluation protocol that reveals the ground-truth risk type to the predictor. Section IV.A explicitly states that predictions were made using only the heatmap corresponding to the annotated risk type, so the reported accuracy is conditioned on the target label by construction. This is a self-definitional evaluation of the claimed risk-presence task. The threshold is unreported, no majority-class baseline or confidence intervals are given, and the Discussion concedes inconsistent performance on fire and trip/fall categories. These issues directly undermine the headline result. Other limitations (undefined phi_distance, unmapped database queries, sparse priors) are correctness/reproducibility concerns rather than circularity. Overall, the method has independent content, but the main reported evaluation is partially circular, warranting a score of 6.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The framework depends on free parameters that are either hand-chosen (k, N) or unreported (threshold, iterations, tolerance, epsilon, and the entire phi_distance kernel), plus domain assumptions about transfer of Japanese accident statistics to NYU V2 scenes and about centroid distance as a proximity proxy. None of these is fitted against an external hold-out set, so the central claim rests on choices that the reader cannot inspect. No new physical or conceptual entity is introduced: phi_accrel and phi_distance are statistics over existing data, and the propagation influence is a scalar weight.

free parameters (5)
  • Laplace smoothing constant k = k = 1
    Hand-chosen in Eqs. (2) and (4) to stabilize low-count ratios; its value directly shapes all risk scores and propagation weights.
  • Binary risk detection threshold = unreported
    Section IV.A thresholds heatmap activation to decide risk present or absent; the value is never stated and appears tuned on the 20 evaluation images.
  • Propagation hyperparameters (max iterations, tolerance, epsilon) = unreported
    Algorithm 1 requires max iterations, a tolerance, and an epsilon for the denominator; none are given, and convergence is asserted via the tolerance without analysis.
  • Spatial distance kernel phi_distance = undefined
    Eq. (6) and Algorithm 1 depend on phi_distance over 3D centroid distances, but its functional form is never specified; any monotone decreasing kernel with different scaling changes propagation.
  • Number of accident types N = N = 3
    Only cut, fire, and trip/fall are modeled; other accident types in the database are dropped, which changes the denominator in Eqs. (2) and (4).
axioms (5)
  • domain assumption Co-occurrence counts in the Japanese accident database reflect genuine accident causation and transfer to camera images from NYU V2.
    Section III.B defines phi_accrel from the database; if the statistical relationships do not transfer across domains (Japan's database to NYU's US scenes), the propagation weights are meaningless.
  • domain assumption Bounding-box centroid distance in 3D approximates the spatial proximity that drives real accidents.
    Section III.B computes phi_distance from centroids of detection boxes; boxes include surrounding space and occlusions, so centroid distance is a crude proxy for contact or reach proximity.
  • ad hoc to paper The iterative propagation with per-iteration min-max renormalization converges and preserves risk ordering.
    Algorithm 1 breaks on a tolerance but no convergence proof is given, and the renormalization changes scores nonlinearly across iterations.
  • standard math Laplace smoothing with k=1 gives stable, unbiased risk ratios.
    Eq. (2) applies Bayesian smoothing; the math is standard, but the choice k=1 and the treatment of N are ad hoc and shift all scores toward uniformity for low-count objects.
  • domain assumption Household accident risk decomposes into the three categories cut, fire, and trip/fall.
    The framework and evaluation are built on exactly three types; other hazard classes (e.g., poison, electric shock) are ignored, which affects both the database denominators and the human-annotation protocol.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Context-Aware Risk Estimation in Home Environments: A Probabilistic Framework for Service Robots." pith.science (2026). https://pith.science/paper/VAQC3PAT

@misc{pith2026250819788,
  author       = {Pith},
  title        = {Pith review of: Context-Aware Risk Estimation in Home Environments: A Probabilistic Framework for Service Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VAQC3PAT}},
  note         = {Machine review of arXiv:2508.19788}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present a novel framework for estimating accident-prone regions in everyday indoor scenes, aimed at improving real-time risk awareness in service robots operating in human-centric environments. As robots become integrated into daily life, particularly in homes, the ability to anticipate and respond to environmental hazards is crucial for ensuring user safety, trust, and effective human-robot interaction. Our approach models object-level risk and context through a semantic graph-based propagation algorithm. Each object is represented as a node with an associated risk score, and risk propagates asymmetrically from high-risk to low-risk objects based on spatial proximity and accident relationship. This enables the robot to infer potential hazards even when they are not explicitly visible or labeled. Designed for interpretability and lightweight onboard deployment, our method is validated on a dataset with human-annotated risk regions, achieving a binary risk detection accuracy of 75%. The system demonstrates strong alignment with human perception, particularly in scenes involving sharp or unstable objects. These results underline the potential of context-aware risk reasoning to enhance robotic scene understanding and proactive safety behaviors in shared human-robot spaces. This framework could serve as a foundation for future systems that make context-driven safety decisions, provide real-time alerts, or autonomously assist users in avoiding or mitigating hazards within home environments.

Figures

Figures reproduced from arXiv: 2508.19788 by Akash Chikhalikar, Ankit A. Ravankar, Jose Victorio Salazar Luces, Sena Ishii, Yasuhisa Hirata.

Figure 1
Figure 1. Figure 1: Overview of the proposed risk propagation framework [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of querying the Accident Information [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of risk propagation in a household environment. The left section represents the initial risk scores based [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of risk score maps before and after risk propagation for different accident types: Cut, Trip/Fall, and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation GUI Flow fire risk score, as depicted in the top row. However, the cabinet in the foreground is also erroneously attributed a high risk. Ideally, only the box and radiator should be highlighted for fire risk in this context. Through risk propagation, shown in the bottom row, the cabinet’s risk score is effectively reduced, and the high fire risk is appropriately highlighted for the box only. Thi… view at source ↗
Figure 6
Figure 6. Figure 6: Risk Detection Confusion Matrix labeled as non-risky. This thresholding step ensures that only strongly activated risk regions influence the binary decision [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Confusion matrices showing risk type prediction [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 22 canonical work pages · 3 internal anchors

  1. [1]

    pi0: A vision- language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. , “ pi0: A vision- language-action flow model for general robot control,” arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Care: Cooperation of ai robot enablers to create a vibrant society,

    A. A. Ravankar, S. A. Tafrishi, J. V . Salazar Luces, F. Seto, and Y . Hirata, “Care: Cooperation of ai robot enablers to create a vibrant society,”IEEE Robotics & Automation Magazine, vol. 30, no. 1, pp. 8– 23, 2023

    work page 2023

  3. [3]

    Enhancing object search in indoor spaces via personalized object- factored ontologies,

    A. Chikhalikar, A. A. Ravankar, J. V . S. Luces, and Y . Hirata, “Enhancing object search in indoor spaces via personalized object- factored ontologies,” 2025

    work page 2025

  4. [4]

    Report on Ac- cident Risk Assessment in Household Environments,

    Ministry of Health, Labour and Welfare of Japan, “Report on Ac- cident Risk Assessment in Household Environments,” Tech. Rep. 0000047617, Ministry of Health, Labour and Welfare of Japan, 2013. [Online; accessed March 22, 2025]

    work page 2013

  5. [5]

    Vital Statistics of Japan, 2023,

    Ministry of Health, Labour and Welfare of Japan, “Vital Statistics of Japan, 2023,” Tech. Rep. 15 all, Ministry of Health, Labour and Welfare of Japan, 2023. [Online; accessed March 22, 2025]

    work page 2023

  6. [6]

    Open vocabulary object search utilizing large language models and fuzzy inferencing,

    A. Chikhalikar, A. A. Ravankar, J. V . S. Luces, and Y . Hirata, “Open vocabulary object search utilizing large language models and fuzzy inferencing,” in 2025 IEEE/SICE International Symposium on System Integration (SII), pp. 345–351, 2025

    work page 2025

  7. [7]

    Physical human–robot interaction,

    S. Haddadin and E. Croft, “Physical human–robot interaction,” Springer Handbook of Robotics , pp. 1835–1874, 2017

    work page 2017

  8. [8]

    Real-time context- aware detection of unsafe events in robot-assisted surgery,

    H. Alemzadeh, V . Raman, C. Li, and K. Glisson, “Real-time context- aware detection of unsafe events in robot-assisted surgery,” inProceed- ings of the 2020 IEEE/IFIP International Conference on Dependable Systems and Networks (DSN) , pp. 25–37, IEEE, 2020

    work page 2020

  9. [9]

    Hazards in Daily Life? Enabling Robots to Proactively Detect and Resolve Anomalies

    Z. Song, G. Ouyang, M. Fang, H. Na, Z. Shi, Z. Chen, Y . Fu, Z. Zhang, S. Jiang, M. Fang, et al., “Hazards in daily life? enabling robots to proactively detect and resolve anomalies,” arXiv preprint arXiv:2411.00781, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    TRG-planner: Traversal Risk Graph-Based Path Planning in Unstructured Environments for Safe and Efficient Navigation

    B. Li, J. Wang, and X. Chen, “Trg-planner: Traversal risk graph- based path planning in unstructured environments,” arXiv preprint arXiv:2501.01806, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  11. [11]

    Robotics-enabled in-home envi- ronment screening for fall risks,

    A. Georgiev and A. Mihailidis, “Robotics-enabled in-home envi- ronment screening for fall risks,” ACM Transactions on Accessible Computing, vol. 14, no. 2, pp. 1–25, 2021

    work page 2021

  12. [12]

    Factor graphs for robot perception,

    F. Dellaert and M. Kaess, “Factor graphs for robot perception,” Foundations and Trends® in Robotics , vol. 6, no. 1-2, pp. 1–139, 2017

    work page 2017

  13. [13]

    Multi-Robot Collaborative Perception with Graph Neural Networks

    Y . Chen, X. Yan, J. Liu, and F.-Y . Wang, “Multi-robot collab- orative perception with graph neural networks,” arXiv preprint arXiv:2201.01760, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  14. [14]

    Semantic-based multi-object search optimization in service robots using probabilistic and contextual priors,

    A. Chikhalikar, A. A. Ravankar, J. Victorio Salazar Luces, and Y . Hirata, “Semantic-based multi-object search optimization in service robots using probabilistic and contextual priors,” IEEE Access, vol. 12, pp. 113151–113164, 2024

    work page 2024

  15. [15]

    Ear- bench: Towards evaluating physical risk awareness for task plan- ning of foundation model-based embodied ai agents,

    Z. Zhu, B. Wu, Z. Zhang, L. Han, Q. Liu, and B. Wu, “Ear- bench: Towards evaluating physical risk awareness for task plan- ning of foundation model-based embodied ai agents,” arXiv preprint arXiv:2408.04449, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Integrating semantic awareness and probabilistic priors for object search in indoor environments,

    A. Chikhalikar, A. A. Ravankar, J. V . S. Luces, and Y . Hirata, “Integrating semantic awareness and probabilistic priors for object search in indoor environments,” in The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) 2023 , 2023

    work page 2023

  17. [17]

    Cooperation of assistive robots to improve productivity in the nursing care field,

    Y . Hirata, J. V . S. Luces, A. A. Ravankar, and S. A. Tafrishi, “Cooperation of assistive robots to improve productivity in the nursing care field,” in The International Symposium of Robotics Research , pp. 287–294, Springer, 2022

    2022

  18. [18]

    Enrichme: Perception and interaction of an assistive robot for the elderly at home,

    S. Cos ¸ar, M. Fernandez-Carmona, R. Agrigoroaie, J. Pages, F. Ferland, F. Zhao, S. Yue, N. Bellotto, and A. Tapus, “Enrichme: Perception and interaction of an assistive robot for the elderly at home,” International Journal of Social Robotics , vol. 12, pp. 779–805, 2020

    work page 2020

  19. [19]

    A novel approach for lidar-based robot localization in a scale-drifted map constructed using monocular slam,

    S. Wang, Y . Kobayashi, A. A. Ravankar, A. Ravankar, and T. Emaru, “A novel approach for lidar-based robot localization in a scale-drifted map constructed using monocular slam,” Sensors, vol. 19, no. 10, p. 2230, 2019

    work page 2019

  20. [20]

    “don’t forget to put the milk back!

    J. F. Mullen, P. Goyal, R. Piramuthu, M. Johnston, D. Manocha, and R. Ghanadan, ““don’t forget to put the milk back!” dataset for enabling embodied agents to detect anomalous situations,” IEEE Robotics and Automation Letters, vol. 9, no. 10, pp. 9087–9094, 2024

    work page 2024

  21. [21]

    Deep-learning-based context-aware multi-level information fusion framework for indoor mobile robots,

    R. Chen, Y . Li, J. Wang, H. Liu, and M. Li, “Deep-learning-based context-aware multi-level information fusion framework for indoor mobile robots,” Sensors, vol. 23, no. 4, p. 996, 2023

    work page 2023

  22. [22]

    Accident Information Data Bank System

    Consumer Affairs Agency, “Accident Information Data Bank System.” https://www.jikojoho.caa.go.jp/ai-national/,

    work page

  23. [23]

    Detecting twenty-thousand classes using image-level supervision,

    X. Zhou, R. Girdhar, A. Joulin, P. Kr ¨ahenb¨uhl, and I. Misra, “Detecting twenty-thousand classes using image-level supervision,” in European conference on computer vision , pp. 350–368, Springer, 2022

    work page 2022

  24. [24]

    Indoor seg- mentation and support inference from rgbd images,

    P. K. Nathan Silberman, Derek Hoiem and R. Fergus, “Indoor seg- mentation and support inference from rgbd images,” in ECCV, 2012

    2012

  25. [25]

    Psm: A predictive safety model for body motion based on the spring-damper pendulum,

    S. A. Tafrishi, A. A. Ravankar, and Y . Hirata, “Psm: A predictive safety model for body motion based on the spring-damper pendulum,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 6657–6664, IEEE, 2022

    work page 2022

  26. [26]

    Safe mobile robot navigation in human-centered en- vironments using a heat map-based path planner,

    A. Ravankar, A. A. Ravankar, Y . Hoshino, M. Watanabe, and Y . Kobayashi, “Safe mobile robot navigation in human-centered en- vironments using a heat map-based path planner,” Artificial Life and Robotics, vol. 25, pp. 264–272, 2020

    2020

  27. [2024]

    11, 2025

    Accessed: Feb. 11, 2025

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.