Embodied Active Learning under Limited Annotation and Navigation Budget for Object Detection

Hadrien Crassous; Joni Pajarinen; Minahil Raza; Mohamed Yassine Kabouri; Riad Akrour

REVIEW 3 major objections 6 minor 54 references

Embodied Active Learning under Limited Annotation and Navigation Budget for Object Detection

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Frame-by-frame prediction disagreement is a better guide than entropy for selecting what a robot should learn to see.

desk verdict Solid incremental embodied AL paper; empirical wins are real but the inconsistency measure's link to true detector errors is unproven. read the letter →

arxiv 2607.15974 v1 pith:CDJ7O4VC submitted 2026-07-17 cs.RO cs.CV

Hadrien Crassous , Mohamed Yassine Kabouri , Minahil Raza , Joni Pajarinen , Riad Akrour This is my paper

classification cs.ROcs.CV

keywords embodiedactivelearningobjectdetectionpredictiondiscrepancyspatialconsistencynavigationbudgetannotationrobotexploration

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

To adapt an object detector to a new environment without wasting human annotations or robot time, the paper proposes an embodied active learning loop: a robot explores, records frames and predictions, and compares consecutive frames in the estimated visual overlap to find places where the detector's own labels disagree. That disagreement—the Prediction Discrepancy—drives both which paths the robot takes and which images get sent to an annotator. The paper claims that under equal navigation and annotation budgets, this self-supervised signal consistently outperforms random selection, object-count difficulty, and prediction entropy, and approaches a ground-truth oracle that cannot be used in practice. Experiments in large simulated home scenes and in a real indoor scene with a quadruped robot support the claim. The value, if true, is that a robot can aim its own retraining data at its failure cases with no extra supervision.

What carries the argument

The central object is the Prediction Discrepancy, Δ, defined for two consecutive frames with an estimated overlapping region as the size of the symmetric difference of the predicted class lists (as multisets) within that region. The overlap is estimated heuristically from the action taken (and, in the real-world experiment, from a homography computed with feature matching and RANSAC). This scalar score serves two roles: as an intrinsic reward/objective for a path-aware Bayesian optimizer that plans robot trajectories via weighted A* with an UCB acquisition, and as the difficulty score in a two-stage sampler that keeps diversity by clustering candidate images with a frozen-encoder feature sim

What would settle it

In a controlled scene, compute the discrepancy using a depth-based geometric overlap (true field-of-view intersection) instead of the action-based heuristic; if the gap between discrepancy-based selection and entropy-based selection vanishes or reverses, the claimed benefit is an artifact of the overlap approximation rather than a genuine property of prediction disagreement.

Watch

Extended reading notes

Core claim

The paper claims that a simple heuristic—count classes detected in one frame but absent (or differently labeled) in the overlapping region of the next frame—identifies exactly the images that, when annotated and added to the training set, most improve an object detector. This 'Prediction Discrepancy' score drives both the navigation policy (through a path-aware Bayesian optimization with an optimistic acquisition) and the sample selector (a two-stage difficulty-plus-diversity sampler). The result is an embodied batch active learning loop that, under fixed annotation and navigation budgets, outperforms random sampling, image difficulty by object count, and prediction entropy in simulated home

Load-bearing premise

The symmetric-difference measure in Eq. (1) is assumed to reflect genuine detector errors rather than harmless viewpoint or occlusion effects, and the overlap region is estimated without depth or geometric alignment.

Editorial extensions

If this is right

Under a fixed combined budget of navigation steps and annotations, the discrepancy-guided agent ends adaptation with a higher mean Average Precision than entropy-, count-, or random-based selection.
The navigation signal and sampling signal can be the same scalar, so the robot's exploration strategy doubles as its data-selection strategy.
The benefit is consistent across three model sizes of the real-time detector and in a real indoor deployment, where the overlap is computed with a homography instead of a heuristic.
The selected samples are measurably harder for the model (lower F1) and more class-balanced (higher class entropy) than baseline selections.
The approach requires no external supervision at selection time; the oracle annotates only after selection.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Depth or geometric alignment could turn the heuristic overlap into a true spatial-consistency signal, potentially strengthening the method and making it robust in cluttered scenes; this is an extension the paper leaves implicit.
The same discrepancy signal could be reused as a reward in a reinforcement-learned exploration policy rather than a hand-designed Bayesian optimizer, merging the curiosity and active-learning lines directly.
The method naturally transfers to other dense vision tasks (e.g., segmentation) where frame-to-frame label disagreement can score informativeness.
A weaker oracle (e.g., no human correction, only foundation-model labels) would stress-test the pipeline end-to-end; the paper uses a semi-automated oracle with human correction.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's my read: the paper is a solid, incremental advance. The contribution is the specific combination of spatial-consistency uncertainty as both a navigation reward and a batch active-learning sample selector under strict navigation and annotation budgets. That integration isn't in the prior work they cite—Chaplot's semantic curiosity needs full trajectory annotations, and Yu et al. use augmentation consistency. The experiments are decently controlled: same budgets across methods, held-out eval, bootstrapped IQM, three YOLO sizes, and a real Spot scene. The PD score beats Entropy, Count, and Random consistently. That's real evidence.

The soft spot is the load-bearing heuristic in Eq. (1). It matches boxes by class only, with no geometric alignment, and the overlap region R(a) is estimated from the action alone. The paper explicitly says it doesn't account for localization errors. That means a harmless viewpoint change—an object leaving the frame, or two different instances of the same class—can produce a nonzero discrepancy that is not a detector failure. Since that same score drives both navigation and sampling, any contamination is amplified. The authors never validate that PD actually correlates with true detector errors. A simple diagnostic—correlating PD with per-frame AP gaps using ground-truth visibility—would settle this. Without it, the empirical wins could be partly due to a spurious correlation with clutter or viewpoint novelty, though the Count baseline mitigates the pure-object-count story.

The other softer spots: they don't run the closest prior method (Chaplot et al.) as a baseline, and the open-source link isn't a pinned artifact. Both are minor in proportion.

The central claim holds up empirically as far as it goes, but the mechanism is under-validated. That's a conditional, not a reject. Who should read this? People working on embodied active learning or active learning for detection. It's a useful combination, not a breakthrough. I'd send it to a serious referee, and I'd cite it if I were working in that area.

My recommendation: engage with it—send it to review, but ask for the diagnostic validation and a comparison with the relevant prior method or a clear argument for why it's out of scope.

Referee Report

3 major / 6 minor

Summary. The paper formulates embodied adaptation of an object detector as batch active learning under both navigation and annotation budgets. The proposed method uses a Prediction Discrepancy score (Eq. 1) — the symmetric difference of class multisets in the estimated overlap region between consecutive frames — to rank candidate images for annotation and to guide an adaptive navigation agent via path-aware Bayesian optimization with a random-forest surrogate and UCB acquisition (Eqs. 2–3). Image selection follows a two-stage PPAL-style difficulty-plus-diversity procedure. Experiments in ProcTHOR with YOLOv5s/m/l compare Discrepancy against Random, Count, Entropy, and an Oracle scoring method, and compare the Adaptive agent against FBE, Sweep, and Random navigation. Real-world validation on a Spot robot with a semi-automated OWLv2 labeling pipeline is also reported. The paper claims that spatial inconsistency improves detector accuracy under the same budgets without external supervision.

Significance. If the central claim holds, the paper offers a practical, geometry-free heuristic for self-supervised informative data collection in embodied active learning, with a plausible mechanism and thorough experimental methodology. Strengths include: controlled budgets across methods, a held-out evaluation set, bootstrap IQM aggregation over 10 scenes × 10 seeds, and a real-robot deployment. The open-source project and the use of standard detectors and simulators also aid reproducibility. The main risk is that the Prediction Discrepancy signal may be contaminated by viewpoint/occlusion effects rather than genuinely reflecting detector errors; because the same signal drives both navigation and sampling, this is load-bearing for the paper's conclusions.

major comments (3)

[§III-B, Eq. (1)] The central claim rests on Prediction Discrepancy being a detector-error signal, but Eq. (1) matches boxes by class only within a heuristically estimated overlap region R(a). Nonzero Δ can arise from benign viewpoint changes, occlusion, or different object instances of the same class, without any detector failure. The manuscript acknowledges this in §III-B ('this measure does not account for localization errors'), but no diagnostic quantifies how much of the observed Δ correlates with genuine errors. Since the same score drives both the navigation surrogate (§III-C) and the annotation sampler (§III-D), this contamination is amplified. I request a validation study: e.g., using the simulator's ground truth, compare Δ for frame pairs with and without per-frame AP degradation, or run an ablation where a geometric/homography-based overlap filter (as used in the real-world setup) is applied in
[§IV-D, Table I and Fig. 3] The claim that Discrepancy selects 'more informative' samples is supported mainly by downstream mAP and by descriptive metrics. However, F1_annot in Table I is computed on the very samples selected by the model; lower F1 is partly a mechanical consequence of selecting hard examples, not independent evidence of informativeness. The Class Distribution Entropy comparison is interesting but descriptive. The Oracle upper-bound comparison is appropriate, but it would strengthen the paper to show per-class AP improvements or an error decomposition (e.g., false negatives vs. false positives) to demonstrate that Discrepancy targets genuine detector failures rather than merely viewpoint variation. The current evidence is suggestive but does not isolate the signal.
[§IV-D, Fig. 5 and §II] The navigation comparison uses only FBE, Sweep, and Random baselines. The closest related work — Chaplot et al. [22], which also uses spatial inconsistency as a reward — is discussed in §II but not compared. Since [22] assumes full trajectory annotation, a direct comparison would require adaptation to the annotation-budget setting, but an ablation that replaces the Discrepancy reward with a generic curiosity signal (e.g., prediction disagreement from an ensemble) would clarify whether the navigation gains come from the specific score or from exploration in general. As it stands, the navigation result supports the overall method but does not isolate the contribution of Eq. (1) to trajectory planning.

minor comments (6)

[§III-B] The multiset symmetric difference in Eq. (1) is nonstandard; please define the matching procedure precisely, especially how duplicate classes are paired and how 'unmatched' elements near borders/small boxes are excluded.
[§III-C] The UCB multiplier 3 and the 50 random destination candidates are heuristic; a sensitivity analysis or a brief justification would be useful, given that these are free parameters.
[§V-A] The real-world experiment uses stored trajectories and scores them offline, rather than the online adaptive navigation of §III-C. This difference should be stated as a limitation in the main text, not only implied in the implementation details.
[Alg. 1] Typo: 'Select a destinations dest' should be 'Select a destination s_dest'.
[Fig. 4] The qualitative examples in Figure 4 are helpful, but the panels for different scoring methods are not matched for scene/state; consider labeling viewpoints or providing quantitative examples of Δ values.
[Table I] In the Real-World columns, the Oracle row is marked '—' because Oracle is excluded; this is fine, but for clarity note that the missing values are intentional, not experimental failures.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the claimed benefit of Prediction Discrepancy is supported by held-out empirical evaluation.

full rationale

The paper's central contribution is an empirically evaluated active-learning selection and navigation heuristic. Eq. (1) defines Prediction Discrepancy as a class-multiset symmetric difference over an estimated overlap region; it is not derived from, nor does it define, the target mAP. The claimed benefit is established by held-out mAP comparisons (Figs. 3, 5, 8) against Random, Count, Entropy, and Oracle, using an independent evaluation set described in Section IV-C. The surrogate navigation model in Section III-C is trained on observed Delta values and used for UCB-based acquisition; this is standard intrinsic-reward feedback, not a prediction of the evaluation metric. Table I's F1_annot metric is a characterization of selected samples, not the central proof, and lower F1 is not guaranteed by Delta by construction, since Delta measures cross-view class disagreement rather than ground-truth error. There are no load-bearing self-citations; Chaplot et al. [22], [38] are cited as related work and differentiated from the proposed setting. The Section III-B caveat that the measure does not account for localization errors is a validity limitation, not circularity: it concerns whether the heuristic measures what it claims, not whether the result is assumed by construction. No equation or fitted parameter reduces to the target result, so the derivation chain is self-contained with respect to circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical or algorithmic entities are introduced; Prediction Discrepancy is a score, not a new object/force/representation. The method relies on one hand-designed heuristic plus standard ML components, with two unablated hand-set constants and an unspecified exclusion threshold in the score.

free parameters (3)

UCB exploration constant β = 3
Used in µ_UCB(s)=µ(s)+3σ(s) in Section III-C for navigation acquisition; set by hand with no sensitivity analysis.
Destination candidate sample size = 50
The acquisition function picks the best among 50 randomly sampled positions (Section III-C); no ablation on this count.
Prediction Discrepancy exclusion thresholds = unspecified
Section III-B excludes unmatched objects that are 'too small or near its border' but values are not reported; this heuristic shapes the score.

assumptions (4)

domain assumption Spatial inconsistency between predictions in overlapping frames is a reliable proxy for object-detection errors and training informativeness
Eq. (1) defines Prediction Discrepancy as this inconsistency; the method's sampling and navigation both depend on this premise (Section III-B).
domain assumption The overlap region R(a) can be estimated from the action alone without depth or image alignment
Section III-B states R(a) 'is estimated heuristically from the action alone'; if the overlap is wrong, matched/mismatched class lists are noisy.
domain assumption Random-forest Bayesian optimization with UCB and path averaging is a sufficient navigation strategy within the time budget
Section III-C adopts BO without a learned policy; effectiveness is assumed from experiments, not proven.
domain assumption The PPAL two-stage pipeline (3×A selection, CCMS diversity clustering with frozen ResNet features) preserves diverse informative samples
Section III-D keeps PPAL's CCMS unchanged; the paper does not re-validate this component for its setting.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Embodied Active Learning under Limited Annotation and Navigation Budget for Object Detection." pith.science (2026). https://pith.science/paper/CDJ7O4VC

@misc{pith2026260715974,
  author       = {Pith},
  title        = {Pith review of: Embodied Active Learning under Limited Annotation and Navigation Budget for Object Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CDJ7O4VC}},
  note         = {Machine review of arXiv:2607.15974}
}

read the original abstract

This paper studies how to adapt a computer vision object detector to an unknown environment under both a robot navigation time and annotation budget constraint. Our approach selects informative robot trajectories and image samples to retrain the detector, explicitly targeting its failure cases. Formally, the approach is an embodied variant of batch active learning, where at each round an agent has a limited navigation budget to collect candidate samples and a limited annotation budget for the most relevant images. We leverage spatial consistency to identify images with inconsistent labels, which are likely to provide the greatest improvement to the vision model. We evaluate the approach using different active learning objectives on large scenes from the AI2-THOR simulator and on a real-world setup using a Boston Dynamics Spot robot with the real-time object detector YOLOv5. Through comparison against several baselines, our experimental results show that spatial inconsistency helps guide the agent and select relevant images without external supervision, achieving the highest detection accuracy at the end of the adaptation process under the same budget. The open-source project can be found at https://mkabouri.github.io/embodied-active-learning-od

Figures

Figures reproduced from arXiv: 2607.15974 by the authors.

**Figure 2.** Examples of Prediction Discrepancy (∆) for two pairs of successive frames. The red area denotes the overlapping region R between the two successive frames. We match predicted objects of the same class across frames and count unmatched objects. (Left) Apple is matched but Pot is unmatched: ∆ = 1. (Right) Both Mug and Book are unmatched: ∆ = 2. The Prediction Discrepancy allows the agent to estimate model prediction e… view at source ↗

**Figure 3.** Aggregate Performance of different scoring methods, across different model sizes (left: small, middle: medium, right: [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Collected data across 3 scoring methods. We visualize [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Aggregate Performance of different navigation meth [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 7.** Figure 7: Objects used for real robot experiments: bottle, box, [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Aggregate Performance of different scoring methods [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 8 linked inside Pith

[22]

Semantic Curiosity for Active Visual Learning: 16th European Conference on Computer Vision, ECCV 2020,

D. S. Chaplot, H. Jiang, S. Gupta, and A. Gupta, “Semantic Curiosity for Active Visual Learning: 16th European Conference on Computer Vision, ECCV 2020,”Computer Vision - ECCV 2020 - 16th European Conference, 2020, Proceedings, pp. 309–326, 2020

2020
[1]

Object Detection With Deep Learning: A Review,

Z.-Q. Zhao, P. Zheng, S.-T. Xu, and X. Wu, “Object Detection With Deep Learning: A Review,”IEEE Transactions on Neural Networks and Learning Systems, vol. 30, pp. 3212–3232, Nov. 2019

2019
[2]

A System for Generalized 3D Multi-Object Search,

K. Zheng, A. Paul, and S. Tellex, “A System for Generalized 3D Multi-Object Search,” inICRA, 2023

2023
[3]

Generalized Object Search,

K. Zheng, “Generalized Object Search,” 2023. arXiv:2301.10121 [cs]

arXiv 2023
[4]

Iqa: Visual question answering in interactive environ- ments,

D. Gordon, A. Kembhavi, M. Rastegari, J. Redmon, D. Fox, and A. Farhadi, “Iqa: Visual question answering in interactive environ- ments,” inCVPR, 2018

2018
[5]

Openeqa: Embodied question answering in the era of foundation models,

A. Majumdar, A. Ajay, X. Zhang, P. Putta, S. Yenamandra, M. Henaff, S. Silwal, P. Mcvay, O. Maksymets, S. Arnaud, K. Yadav, Q. Li, B. Newman, M. Sharma, V . Berges, S. Zhang, P. Agrawal, Y . Bisk, D. Batra, M. Kalakrishnan, F. Meier, C. Paxton, A. Sax, and A. Ra- jeswaran, “Openeqa: Embodied question answering in the era of foundation models,” inCVPR, 2024

2024
[6]

Behavior- 1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation,

C. Li, R. Zhang, J. Wong, C. Gokmen, S. Srivastava, R. Mart ´ın-Mart´ın, C. Wang, G. Levine, W. Ai, B. Martinez, H. Yin, M. Lingelbach, M. Hwang, A. Hiranaka, S. Garlanka, A. Aydin, S. Lee, J. Sun, M. Anvari, M. Sharma, D. Bansal, S. Hunter, K.-Y . Kim, A. Lou, C. R. Matthews, I. Villa-Renteria, J. H. Tang, C. Tang, F. Xia, Y . Li, S. Savarese, H. Gweon, ...

2024
[7]

Embodied Instruction Following in Unknown Environments,

Z. Wu, Z. Wang, X. Xu, H. Yin, Y . Liang, A. Ma, J. Lu, and H. Yan, “Embodied Instruction Following in Unknown Environments,” inIROS, 2025. ISSN: 2153-0866

2025

Show all 54 references

[8]

Embedded solution to detect and classify head level objects using stereo vision for visually impaired people with audio feedback,

K. Mu ˜noz, M. Chavarria, L. Ortiz, S. Suter, K. Sch ¨onenberger, and B. Bacca-Cortes, “Embedded solution to detect and classify head level objects using stereo vision for visually impaired people with audio feedback,”Scientific Reports, no. 1, 2025

2025
[9]

Education robot object detection with a brain-inspired approach integrating Faster R-CNN, YOLOv3, and semi-supervised learning,

Q. Hong, H. Dong, W. Deng, and Y . Ping, “Education robot object detection with a brain-inspired approach integrating Faster R-CNN, YOLOv3, and semi-supervised learning,”Frontiers in Neurorobotics, 2024

2024
[10]

Towards Optimal Correlational Object Search,

K. Zheng, R. Chitnis, Y . Sung, G. Konidaris, and S. Tellex, “Towards Optimal Correlational Object Search,” inICRA, 2022

2022
[11]

You only look once: Unified, real-time object detection,

J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” inCVPR, 2016

2016
[12]

ultralytics/yolov5: v7.0 - YOLOv5 SOTA Realtime Instance Segmentation,

G. Jocher, A. Chaurasia, A. Stoken, J. Borovec, NanoCode012, Y . Kwon, K. Michael, TaoXie, J. Fang, imyhxy, Lorna, Z. Yifu, C. Wong, A. V , D. Montes, Z. Wang, C. Fati, J. Nadar, Laughing, UnglvKitDe, V . Sonck, tkianai, yxNONG, P. Skalski, A. Hogan, D. Nair, M. Strobel, and M...

2022
[13]

DETRs Beat YOLOs on Real-time Object Detection,

Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “DETRs Beat YOLOs on Real-time Object Detection,” 2024. arXiv:2304.08069 [cs]

2024 arXiv
[14]

Microsoft COCO: Common Objects in Context,

T.-Y . Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Doll´ar, and C. L. Zitnick, “Microsoft COCO: Common Objects in Context,” inECCV, 2014

2014
[15]

The Pascal Visual Object Classes (VOC) Challenge,

M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, “The Pascal Visual Object Classes (VOC) Challenge,” International Journal of Computer Vision, vol. 88, pp. 303–338, June 2010

2010
[16]

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,” inECCV, Springer Nature Switzerland, 2022

2022
[17]

Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection,

S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, J. Zhu, and L. Zhang, “Grounding DINO: Marrying DINO with Grounded Pre-training for Open-Set Object Detection,” in ECCV, 2025

2025
[18]

Scaling vision transformers,

X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer, “Scaling vision transformers,” inCVPR, 2022

2022
[19]

A robust learning approach to domain adaptive object detection,

M. Khodabandeh, A. Vahdat, M. Ranjbar, and W. G. Macready, “A robust learning approach to domain adaptive object detection,” in ICCV, 2019

2019
[20]

Domain adaptive faster r-cnn for object detection in the wild,

Y . Chen, W. Li, C. Sakaridis, D. Dai, and L. Van Gool, “Domain adaptive faster r-cnn for object detection in the wild,” inCVPR, 2018

2018
[21]

Interactive data collection for deep learning object detectors on humanoid robots,

E. Maiettini, G. Pasquale, L. Rosasco, and L. Natale, “Interactive data collection for deep learning object detectors on humanoid robots,” in 2017 IEEE-RAS 17th International Conference on Humanoid Robotics (Humanoids), 2017. ISSN: 2164-0580

2017
[23]

So you think you can scale up autonomous robot data collection?,

S. Mirchandani, S. Belkhale, J. Hejna, E. Choi, M. S. Islam, and D. Sadigh, “So you think you can scale up autonomous robot data collection?,” inConference on Robot Learning (CoRL), Proceedings of Machine Learning Research, PMLR, 2025

2025
[24]

Embodied Domain Adapta- tion for Object Detection,

X. Shi, Y . Qiao, L. Liu, and F. Dayoub, “Embodied Domain Adapta- tion for Object Detection,” inIROS, 2025. ISSN: 2153-0866

2025
[25]

AI2-THOR: An Interactive 3D Environment for Visual AI,

E. Kolve, R. Mottaghi, W. Han, E. VanderBilt, L. Weihs, A. Herrasti, M. Deitke, K. Ehsani, D. Gordon, Y . Zhu, A. Kembhavi, A. Gupta, and A. Farhadi, “AI2-THOR: An Interactive 3D Environment for Visual AI,” 2022. arXiv:1712.05474 [cs]

2022 arXiv
[26]

ProcTHOR: Large-Scale Embodied AI Using Procedural Generation,

M. Deitke, E. VanderBilt, A. Herrasti, L. Weihs, J. Salvador, K. Ehsani, W. Han, E. Kolve, A. Farhadi, A. Kembhavi, and R. Mottaghi, “ProcTHOR: Large-Scale Embodied AI Using Procedural Generation,”
[27]

Rich feature hierarchies for accurate object detection and semantic segmentation,

R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation,” inCVPR, 2014

2014
[28]

Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Networks,” inAdvances in Neural Information Processing Systems, Curran Associates, Inc., 2015

2015
[29]

SSD: Single Shot MultiBox Detector,

W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y . Fu, and A. C. Berg, “SSD: Single Shot MultiBox Detector,” inECCV, 2016

2016
[30]

YOLOv3: An Incremental Improvement,

J. Redmon and A. Farhadi, “YOLOv3: An Incremental Improvement,”
[31]

End-to-End Object Detection with Transformers,

N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-End Object Detection with Transformers,” in ECCV, Springer International Publishing, 2020

2020
[32]

Multi-class active learning for image classification,

A. J. Joshi, F. Porikli, and N. Papanikolopoulos, “Multi-class active learning for image classification,” inCVPR, 2009. ISSN: 1063-6919

2009
[33]

Query-Based Hard-Image Retrieval for Object Detection at Test Time,

E. Ayers, J. Sadeghi, J. Redford, R. Mueller, and P. K. Dokania, “Query-Based Hard-Image Retrieval for Object Detection at Test Time,”AAAI, no. 12, 2023

2023
[34]

Heterogeneous Uncertainty Sampling for Supervised Learning,

D. D. Lewis and J. Catlett, “Heterogeneous Uncertainty Sampling for Supervised Learning,” inMachine Learning Proceedings 1994, Elsevier, 1994

1994
[35]

Learning loss for active learning,

D. Yoo and I. S. Kweon, “Learning loss for active learning,” inCVPR, 2019

2019
[36]

Plug and play active learning for object detection,

C. Yang, L. Huang, and E. J. Crowley, “Plug and play active learning for object detection,” inCVPR, 2024

2024
[37]

Consistency-based active learning for object detection,

W. Yu, S. Zhu, T. Yang, and C. Chen, “Consistency-based active learning for object detection,” inCVPR Workshops, 2022

2022
[38]

SEAL: Self-supervised Embodied Active Learning using Exploration and 3D Consistency,

D. S. Chaplot, M. Dalal, S. Gupta, J. Malik, and R. R. Salakhutdinov, “SEAL: Self-supervised Embodied Active Learning using Exploration and 3D Consistency,” inAdvances in Neural Information Processing Systems, Curran Associates, Inc., 2021

2021
[39]

A frontier-based approach for autonomous exploration,

B. Yamauchi, “A frontier-based approach for autonomous exploration,” inProceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation CIRA’97. ’Towards New Computational Principles for Robotics and Automation’, 1997

1997
[40]

R. S. Sutton and A. G. Barto,Reinforcement Learning: An Introduc- tion. The MIT Press, second ed., 2018

2018
[41]

What is intrinsic motivation? A typol- ogy of computational approaches,

P.-Y . Oudeyer and F. Kaplan, “What is intrinsic motivation? A typol- ogy of computational approaches,”Frontiers in Neurorobotics, vol. 1, Nov. 2007

2007
[42]

Curiosity-driven Exploration by Self-supervised Prediction,

D. Pathak, P. Agrawal, A. A. Efros, and T. Darrell, “Curiosity-driven Exploration by Self-supervised Prediction,” inProceedings of the 34th International Conference on Machine Learning, pp. 2778–2787, PMLR, July 2017

2017
[43]

Exploration by Random Network Distillation,

Y . Burda, H. Edwards, A. Storkey, and O. Klimov, “Exploration by Random Network Distillation,” Oct. 2018. arXiv:1810.12894 [cs]

2018 arXiv
[44]

AutoNeRF: Training Implicit Scene Representations with Autonomous Agents,

P. Marza, L. Matignon, O. Simonin, D. Batra, C. Wolf, and D. S. Chaplot, “AutoNeRF: Training Implicit Scene Representations with Autonomous Agents,” inIROS, 2024. ISSN: 2153-0866

2024
[45]

Proximal Policy Optimization Algorithms,

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal Policy Optimization Algorithms,” 2017. arXiv:1707.06347 [cs]

2017 arXiv
[46]

Path-aware optimistic optimization for a mobile robot,

T. S ˆantejudean and L. Bus ¸oniu, “Path-aware optimistic optimization for a mobile robot,” in2021 60th IEEE Conference on Decision and Control (CDC), 2021. ISSN: 2576-2370

2021
[47]

A Tutorial on Bayesian Optimization,

P. I. Frazier, “A Tutorial on Bayesian Optimization,” 2018. arXiv:1807.02811 [stat]

2018 arXiv
[48]

Random Forests,

L. Breiman, “Random Forests,”Machine Learning, vol. 45, pp. 5–32, Oct. 2001

2001
[49]

Finite-time analysis of the multiarmed bandit problem,

P. Auer, N. Cesa-Bianchi, and P. Fischer, “Finite-time analysis of the multiarmed bandit problem,”Mach. Learn., vol. 47, p. 235–256, May 2002

2002
[50]

Deep reinforcement learning at the edge of the statis- tical precipice,

R. Agarwal, M. Schwarzer, P. S. Castro, A. C. Courville, and M. Bellemare, “Deep reinforcement learning at the edge of the statis- tical precipice,”Advances in Neural Information Processing Systems, vol. 34, 2021

2021
[51]

Spot® — Boston Dynamics,

Boston Dynamics, “Spot® — Boston Dynamics,” 2024. [Online]

2024
[52]

Scaling Open- V ocabulary Object Detection,

M. Minderer, A. Gritsenko, and N. Houlsby, “Scaling Open- V ocabulary Object Detection,”Advances in Neural Information Pro- cessing Systems, vol. 36, pp. 72983–73007, Dec. 2023

2023
[2018]

arXiv:1804.02767 [cs]

arXiv
[2022]

arXiv:2206.06994 [cs]

arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [22]

Semantic Curiosity for Active Visual Learning: 16th European Conference on Computer Vision, ECCV 2020,

D. S. Chaplot, H. Jiang, S. Gupta, and A. Gupta, “Semantic Curiosity for Active Visual Learning: 16th European Conference on Computer Vision, ECCV 2020,”Computer Vision - ECCV 2020 - 16th European Conference, 2020, Proceedings, pp. 309–326, 2020

2020

[2] [1]

Object Detection With Deep Learning: A Review,

Z.-Q. Zhao, P. Zheng, S.-T. Xu, and X. Wu, “Object Detection With Deep Learning: A Review,”IEEE Transactions on Neural Networks and Learning Systems, vol. 30, pp. 3212–3232, Nov. 2019

2019

[3] [2]

A System for Generalized 3D Multi-Object Search,

K. Zheng, A. Paul, and S. Tellex, “A System for Generalized 3D Multi-Object Search,” inICRA, 2023

2023

[4] [3]

Generalized Object Search,

K. Zheng, “Generalized Object Search,” 2023. arXiv:2301.10121 [cs]

arXiv 2023

[5] [4]

Iqa: Visual question answering in interactive environ- ments,

D. Gordon, A. Kembhavi, M. Rastegari, J. Redmon, D. Fox, and A. Farhadi, “Iqa: Visual question answering in interactive environ- ments,” inCVPR, 2018

2018

[6] [5]

Openeqa: Embodied question answering in the era of foundation models,

A. Majumdar, A. Ajay, X. Zhang, P. Putta, S. Yenamandra, M. Henaff, S. Silwal, P. Mcvay, O. Maksymets, S. Arnaud, K. Yadav, Q. Li, B. Newman, M. Sharma, V . Berges, S. Zhang, P. Agrawal, Y . Bisk, D. Batra, M. Kalakrishnan, F. Meier, C. Paxton, A. Sax, and A. Ra- jeswaran, “Openeqa: Embodied question answering in the era of foundation models,” inCVPR, 2024

2024

[7] [6]

Behavior- 1k: A human-centered, embodied ai benchmark with 1,000 everyday activities and realistic simulation,

C. Li, R. Zhang, J. Wong, C. Gokmen, S. Srivastava, R. Mart ´ın-Mart´ın, C. Wang, G. Levine, W. Ai, B. Martinez, H. Yin, M. Lingelbach, M. Hwang, A. Hiranaka, S. Garlanka, A. Aydin, S. Lee, J. Sun, M. Anvari, M. Sharma, D. Bansal, S. Hunter, K.-Y . Kim, A. Lou, C. R. Matthews, I. Villa-Renteria, J. H. Tang, C. Tang, F. Xia, Y . Li, S. Savarese, H. Gweon, ...

2024

[8] [7]

Embodied Instruction Following in Unknown Environments,

Z. Wu, Z. Wang, X. Xu, H. Yin, Y . Liang, A. Ma, J. Lu, and H. Yan, “Embodied Instruction Following in Unknown Environments,” inIROS, 2025. ISSN: 2153-0866

2025