Pith. sign in

REVIEW 5 major objections 6 minor 35 references

A Hybrid Mamba for Audio-Visual Navigation

T0 review · 5 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

Pith's one-line read Replacing GRU and CNN backbones with Mamba selective state-space models lifts audio-visual navigation success rates by 11.3 points on Matterport3D and roughly 20 points on Replica.

desk verdict Sensible Mamba-for-AVN idea, but the current empirical write-up doesn't back the headline gains. read the letter →

arxiv 2607.13110 v1 pith:2TY7V4JW submitted 2026-07-14 cs.LG cs.AIeess.SP

classification cs.LGcs.AIeess.SP
keywords audio-visualnavigationstate-spacemodelsMambaselectivescanaudioencoderreinforcementlearninggeneralizationembodiedAI
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

This paper argues that the five-year-old backbone of audio-visual navigation—CNN audio encoders plus GRU state trackers—is the main bottleneck, and that replacing both with Mamba-style selective state-space models yields large gains on unheard sounds and unseen scenes. On Matterport3D the proposed Samba architecture improves success rate by 11.3 percentage points over the prior state-of-the-art, and by roughly 20 points on Replica, while using fewer parameters. The authors trace the gains to two components: a bidirectional Audio Mamba Encoder that captures global time-frequency structure in spectrograms, and a Mamba State Encoder whose input-dependent selection prevents acoustic landmarks from being diluted in the hidden state.

What carries the argument

Two components carry the argument: the Mamba State Encoder (M-SE), a discretized continuous state-space model whose A, B, C, Δ parameters are generated from the input, enabling content-aware memory updates; and the bidirectional Audio Mamba Encoder (AME), which flattens spectrogram channels and frequencies into a token sequence, alternates forward and backward selective scans across Mamba layers, and pools to an audio state vector. The former replaces the GRU, the latter replaces the CNN audio encoder; the visual branch stays convolutional to keep spatial inductive bias.

What would settle it

Retrain the previous state-of-the-art baseline and Samba under identical conditions (same seed budget, same PPO hyperparameters, same split) and compare unheard-scene success rates across multiple seeds; if the gap falls within a few points or reverses when the baseline's hidden state is given the same sequence length as M-SE, the central claim is not supported.

Watch

Extended reading notes

Core claim

Samba is a hybrid architecture for audio-visual navigation that swaps the two standard backbone modules for state-space counterparts. The Audio Mamba Encoder tokenizes the binaural spectrogram and scans it forward and backward with selective Mamba layers, so interaural time and level differences are modeled globally rather than locally. The Mamba State Encoder replaces the GRU's fixed gating with input-conditioned transition matrices, letting the hidden state retain critical acoustic cues while ignoring redundant observations. The paper claims that on the unheard/unseen test setting this raises success rate from 52.8% to 72.8% on Replica and from 56.7% to 68.0% on MP3D, with parameter counts

Load-bearing premise

The evaluation protocol—the fixed 78/11/18 sound split, unseen test scenes, and the SoundSpaces simulator—is taken as a fair test of generalization, and no seeds, code, or training-variance numbers are given; if the split or undisclosed training details favor the new architecture, the headline gains would not transfer.

Editorial extensions

If this is right

  • If the claimed gains hold, Mamba-based backbones become the new state of the art for audio-visual navigation, with larger margins on generalization settings (unheard sounds, unseen scenes) than on familiar ones.
  • The same backbone substitution could extend to other embodied navigation tasks that use recurrent state trackers and convolutional encoders.
  • The reported parameter reductions—about 30% in the state encoder and 43% in the audio encoder—suggest selective state-space models can match or beat recurrent and convolutional accuracy at lower cost, which matters for on-device robotics.
  • The dual parallel/recurrent execution mode of M-SE means the same architecture can train with long-sequence parallelism and deploy in single-step mode, a practical advantage for real-time agents.

Reading between the lines

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

  • If the architecture is the main driver, the same two-module swap may also benefit audio-visual grounding or sound-source localization tasks that share the spectrogram-to-state structure, even outside navigation.
  • The paper's own ablations show that removing the audio encoder causes a larger drop on Replica than removing the state encoder; a natural follow-up is isolating whether AME's global time-frequency modeling is the key factor, and whether a bidirectional GRU would capture some of the same benefit.
  • A stronger test would be reporting variance across seeds and ablating with matched training budgets; until then, the 11–20 point gains should be treated as conditional on the specific simulator split.
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.

Referee Report

5 major / 6 minor

Summary. The paper proposes Samba, an audio-visual navigation architecture built on Mamba/SSM modules. The Audio Mamba Encoder (AME) replaces a CNN audio encoder for binaural spectrograms, and the Mamba State Encoder (M-SE) replaces a GRU temporal-state encoder. The authors evaluate on SoundSpaces with Replica and MP3D, under heard/unheard sound and seen/unseen scene conditions, reporting SR, SPL, and SNA. Their headline result is that in the unheard/unseen condition Samba raises SR from AV-WaN's 52.8% to 72.8% on Replica and from 56.7% to 68.0% on MP3D, with lower parameter counts (Table II). Ablations isolate the contributions of AME and M-SE.

Significance. If substantiated, the paper would establish that a selective state-space backbone is a viable and efficient replacement for the CNN+GRU backbone that has dominated audio-visual navigation since SoundSpaces. The specific technical idea—using a Mamba layer for temporal state encoding and a bidirectional Mamba scan over spectrogram tokens—is plausible and worth investigating. The paper does not provide code, seeds, variance estimates, or full training details, and its internal ablation tables are inconsistent with the text (e.g., the w/o M-SE variant outperforms the full model on the Replica Heard condition). The empirical foundation is therefore not yet adequate to support the claimed level of gains or the attribution to M-SE.

major comments (5)
  1. [§IV, Table I] No measure of variability or reproducibility is reported. Table I lists single numbers for SPL/SR/SNA, with no standard deviations, number of seeds, test episode counts, or PPO hyperparameters. For RL navigation, the difference between Samba and AV-WaN on MP3D Heard (SR 95.0 vs 93.6, SPL 73.3 vs 72.3) is within typical seed noise. Without this information, the +11.3 pp claim cannot be evaluated. Please report seed-level results, episode counts, hyperparameters, and ideally release code.
  2. [§IV.B, Table I] The text says 'Samba consistently outperforms all existing baselines across all test environments,' but Table I shows AV-WaN outperforms Samba on Replica Heard by a large margin: AV-WaN SPL 86.6 vs 78.3, SR 98.7 vs 93.4, SNA 70.7 vs 63.3. This contradicts the 'state-of-the-art' summary and the 'consistently outperforms' claim. The claim needs to be restricted or the discrepancy explained.
  3. [§IV.D, Tables III & IV] Table III is titled 'Ablation study on the Fusion Controller of CRFN,' not on Samba; this appears to be a copy-paste artifact. More substantively, the w/o M-SE row in Table III beats the full model on Replica Heard in all metrics (SPL 82.5 vs 78.3, SR 97.3 vs 93.4, SNA 66.5 vs 63.3). Thus the M-SE component degrades performance on the heard setting, and the paper's framing of M-SE as a core contribution is not supported by the ablation evidence. The text should report and discuss this regression, and the attribution of gains to the hybrid architecture must be recalibrated.
  4. [§IV Experimental Setup] The evaluation protocol is not anchored to an external standard. The 78/11/18 sound split and the 'all test scenes are unseen during training' rule are attributed only to the authors' earlier preprint [15]. It is unclear whether the baseline numbers in Table I were obtained under the same split, the same episode distribution, and the same number of seeds. If not, the comparison is confounded. Please specify the exact provenance of each baseline row and justify the split relative to the SoundSpaces/AVWaN protocol.
  5. [§III.C] The M-SE section does not give the equations for the input-dependent discretization parameters Δ, B, C, nor the exact state-reset/masking logic beyond a verbal description. Since the architecture claim depends on this selective mechanism, the method is not fully reproducible as written. Please provide the formal definitions or pseudocode for M-SE, matching the detail given for AME in Algorithm 1.
minor comments (6)
  1. [Fig. 1] The figure contains placeholder text '?? = Δ?, Δ?' and is low resolution; please redraw.
  2. [References] Reference [4] is used for 'structured state-space model,' 'S4,' 'Mamba,' and 'selective scan' interchangeably; please use separate citations.
  3. [Eq. (4)] Equation (4) states the ZOH formula for \bar B, then immediately replaces it by \Delta B; clarify which approximation is actually implemented.
  4. [Table II] Table II's column headers are ambiguous ('Total (M)↓ Replica MP3D'); separate dataset-specific totals in subcolumns.
  5. [Tables III–IV] The term 'Fusion Controller of CRFN' appears in the titles but CRFN is not introduced in the method; remove or explain.
  6. [§IV Experimental Setup] The paper says '102 distinct natural sounds' and the Heard condition is a telephone ringtone; specify whether the ringtone is among the 102 and how many test episodes per sound/scene are used.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical gains are measured on held-out unheard sounds and unseen scenes; no result reduces to its inputs by construction.

full rationale

Samba's central claim is an empirical architecture comparison (Table I) plus ablations (Tables III–IV). No equation in the paper derives the reported SR/SPL/SNA from a fitted constant; Eqs. (1)–(5) are standard tokenization, SSM discretization, and recurrence definitions imported from Mamba [4], and the reported numbers come from SoundSpaces/Habitat evaluation, not from those equations. The only self-referential element is that the 78/11/18 unheard/unseen split is attributed to the authors' own CRFN paper [15]. That is a benchmark-provenance and reproducibility concern (the split is not the original AV-WaN protocol, and no code/seeds/error bars are given), but it is not circularity: the split does not by construction make Samba's success rate high, and the architecture was not fit to the test-set numbers reported. Related-work self-citations [15]–[27] are not load-bearing for the architectural derivation. The internal contradiction in Table I (Samba is worse than AV-WaN on Replica Heard despite the text's 'consistently outperforms' claim) and the mislabeled ablation captions mentioning CRFN are correctness/quality issues, not circular reductions. Hence the derivation chain is self-contained with respect to circularity, and the score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central gains rest on learned SSM quantities (Δ, A, B, C) whose parameterization is unspecified, on omitted PPO hyperparameters, on SoundSpaces' fidelity as a proxy for real acoustic navigation, and on a sound/scene split attributed to the authors' own CRFN paper. No new physical entities are introduced; AME/M-SE are neural modules, not postulates requiring independent evidence.

free parameters (4)
  • Discretization step Δ (M-SE/AME) = not reported
    Input-dependent learned step size; controls memory decay in the selective SSM and is central to the claimed adaptive state tracking.
  • SSM matrices A, B, C (M-SE/AME) = not reported
    Core state-space parameters; no initialization, parameterization, or selective-generation formulas are provided.
  • Number of Mamba layers N_layers in AME = unspecified
    Algorithm 1 loops over N_layers; depth is chosen by hand and not reported.
  • PPO training hyperparameters = unspecified
    Learning rate, horizon, entropy coefficient, number of steps, and environment details are omitted; these strongly affect SR/SPL and the comparison to baselines.
assumptions (5)
  • standard math Zero-order hold discretization of the SSM is valid; the paper then simplifies to bar B = ΔB (Euler-like update).
    Used in Section III-C, Eqs. 3–5. It assumes A and B are constant during each step; the paper does not specify which discretization is actually used in training.
  • domain assumption Flattening the binaural spectrogram into per-time-step tokens preserves interaural time and level difference cues.
    Eq. (1) flattens channel×frequency into a vector per time step; this presupposes a linear projection can retain the localization cues.
  • domain assumption SoundSpaces simulation is a faithful proxy for real audio-visual navigation.
    All results are in simulation; no real-robot validation is provided, and the claims are framed broadly as 'embodied representation' gains.
  • domain assumption PPO can train the recurrent selective-SSM state without instability; episode-end state reset suffices.
    Section III-C introduces masking/reset logic, but no analysis or experiment demonstrates stable RL training with Mamba, leaving open a possible source of the ablation inconsistencies.
  • ad hoc to paper The 78/11/18 sound split and unseen-scene rule defined in [15] is a standard, unbiased benchmark.
    The split is attributed to the authors' own CRFN paper [15] instead of the original SoundSpaces/AVWaN protocol, making the benchmark partially self-referential.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hybrid Mamba for Audio-Visual Navigation." pith.science (2026). https://pith.science/paper/2TY7V4JW

@misc{pith2026260713110,
  author       = {Pith},
  title        = {Pith review of: A Hybrid Mamba for Audio-Visual Navigation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2TY7V4JW}},
  note         = {Machine review of arXiv:2607.13110}
}
read the original abstract

Since the paradigm centered on convolutional neural networks and recurrent architectures was established in 2020, the fundamental backbone networks for audio-visual navigation have undergone no essential changes for more than five years, making them inadequate to support efficient representation of dynamic multimodal sequences. This paper proposes Samba(A Hybrid Mamba for Audio-Visual Navigation). It uses the adaptive selection-enabled Mamba State Encoder (M-SE) to replace conventional GRUs for temporal aggregation, and constructs an Audio Mamba Encoder (AME) to remedy the limitations of convolutional operators in capturing global time-frequency dependencies in spectrograms. Experiments demonstrate that Samba exhibits exceptional generalization performance when facing unheard sound sources and unseen scenes. On the Matterport3D dataset, it improves the navigation success rate (SR) by 11.3\% compared with existing state-of-the-art models, and the performance gain is even more pronounced on the Replica dataset, which features finer scene structures. Such modernized architectural reconstruction unlocks stronger embodied representation capabilities at a lower computational cost, thereby providing a highly robust technical pathway for paradigm evolution in the field of audio-visual navigation.

Figures

Figures reproduced from arXiv: 2607.13110 by the authors.

Figure 1
Figure 1. Overall Architecture of the Samba Framework. fundamental level [18], [19], [20], [21], [22], [23], [24], [25], [26], [27]. B. State Space Models and Mamba State Space Models (SSMs) [4], as an emerging deep learning architecture, are gradually becoming a strong alter￾native to Recurrent Neural Networks (RNNs) [28] and Trans￾formers [29] due to their superior performance in handling extremely long-sequence tasks. From… view at source ↗
Figure 2
Figure 2. Architectural Comparison between the baseline GRU and our proposed Mamba State Encoder (M-SE). The left side illustrates the classic Gated Recurrent Unit, while the right side details the Audio Mamba State Encoder designed in this paper. local receptive fields in traditional convolutional architectures when processing complex acoustic observations, we design an Audio Mamba Encoder (AME) based on selective state￾spac… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 3 linked inside Pith

  1. [15]

    Residual cross-modal fusion networks for audio-visual navigation,

    Y . Wang, Y . Yu, and B. Ren, “Residual cross-modal fusion networks for audio-visual navigation,”arXiv preprint arXiv:2601.08868, 2026

  2. [1]

    Soundspaces: Audio-visual navigation in 3d environments,

    C. Chen, U. Jain, C. Schissler, S. V . A. Gari, Z. Al-Halah, V . K. Ithapu, P. Robinson, and K. Grauman, “Soundspaces: Audio-visual navigation in 3d environments,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF, 2020, pp. 17–36

  3. [2]

    Look, listen, and act: Towards audio-visual embodied navigation,

    C. Gan, Y . Zhang, J. Wu, B. Gong, and J. B. Tenenbaum, “Look, listen, and act: Towards audio-visual embodied navigation,” in2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 9701–9707

  4. [3]

    Learning to set waypoints for audio-visual navigation,

    C. Chen, S. Majumder, Z. Al-Halah, R. Gao, S. K. Ramakrishnan, and K. Grauman, “Learning to set waypoints for audio-visual navigation,” in9th International Conference on Learning Representations, ICLR

  5. [4]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inFirst conference on language modeling, 2024

  6. [5]

    A survey on sound source localization in robotics: From binaural to array processing methods,

    S. Argentieri, P. Danes, and P. Sou `eres, “A survey on sound source localization in robotics: From binaural to array processing methods,” Computer Speech & Language, vol. 34, no. 1, pp. 87–112, 2015

  7. [6]

    Echo-enhanced embodied visual navigation,

    Y . Yu, L. Cao, F. Sun, C. Yang, H. Lai, and W. Huang, “Echo-enhanced embodied visual navigation,”Neural Computation, vol. 35, no. 5, pp. 958–976, 04 2023

  8. [7]

    Dope: Dual object perception-enhancement network for vision-and-language navigation,

    Y . Yu and D. Yang, “Dope: Dual object perception-enhancement network for vision-and-language navigation,” inProceedings of the 2025 International Conference on Multimedia Retrieval, 2025, pp. 1739–1748

Show all 35 references
  1. [8]

    Sound adversarial audio-visual navigation,

    Y . Yu, W. Huang, F. Sun, C. Chen, Y . Wang, and X. Liu, “Sound adversarial audio-visual navigation,” inThe Tenth International Con- ference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022

  2. [9]

    Caven: An embodied conversational agent for efficient audio-visual navigation in noisy environments,

    X. Liu, S. Paul, M. Chatterjee, and A. Cherian, “Caven: An embodied conversational agent for efficient audio-visual navigation in noisy environments,” inProceedings of the AAAI conference on artificial intelligence, vol. 38, 2024, pp. 3765–3773

  3. [10]

    Omnidirectional information gathering for knowledge transfer-based audio-visual navi- gation,

    J. Chen, W. Wang, S. Liu, H. Li, and Y . Yang, “Omnidirectional information gathering for knowledge transfer-based audio-visual navi- gation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 993–11 003

  4. [11]

    Ai2-thor: An interac- tive 3d environment for visual ai,

    E. Kolve, R. Mottaghi, W. Han, E. VanderBilt, L. Weihs, A. Herrasti, M. Deitke, K. Ehsani, D. Gordon, Y . Zhuet al., “Ai2-thor: An interac- tive 3d environment for visual ai,”arXiv preprint arXiv:1712.05474, 2017

  5. [12]

    Measuring acoustics with collaborative multiple agents,

    Y . Yu, C. Chen, L. Cao, F. Yang, and F. Sun, “Measuring acoustics with collaborative multiple agents,” inProceedings of the 32nd Inter- national Joint Conference on Artificial Intelligence (IJCAI-23), Macao, China, August 2023

  6. [13]

    Habitat: A platform for embodied ai research,

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Maliket al., “Habitat: A platform for embodied ai research,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 9339–9347

  7. [14]

    Habitat 2.0: Training home assistants to rearrange their habitat,

    A. Szot, A. Clegg, E. Undersander, E. Wijmans, Y . Zhao, J. Turner, N. Maestre, M. Mukadam, D. S. Chaplot, O. Maksymetset al., “Habitat 2.0: Training home assistants to rearrange their habitat,” Advances in neural information processing systems, vol. 34, pp. 251– 266, 2021

  8. [16]

    Pay self-attention to audio-visual navigation,

    Y . Yu, L. Cao, F. Sun, X. Liu, and L. Wang, “Pay self-attention to audio-visual navigation,” in33rd British Machine Vision Conference 2022, BMVC 2022. London, UK: BMV A Press, November 2022, p. 46

  9. [17]

    Audio-guided dynamic modality fusion with stereo-aware attention for audio-visual naviga- tion,

    J. Li, Y . Yu, L. Wang, F. Sun, and W. Zheng, “Audio-guided dynamic modality fusion with stereo-aware attention for audio-visual naviga- tion,” inInternational Conference on Neural Information Processing. Springer, 2025, pp. 346–359

  10. [18]

    Dgfnet: End-to-end audio-visual source separation based on dynamic gating fusion,

    Y . Yu and S. Sun, “Dgfnet: End-to-end audio-visual source separation based on dynamic gating fusion,” inProceedings of the 2025 Interna- tional Conference on Multimedia Retrieval, 2025, pp. 1730–1738

  11. [19]

    Dynamic multi-target fusion for efficient audio-visual navigation,

    Y . Yu, H. Zhang, and M. Zhu, “Dynamic multi-target fusion for efficient audio-visual navigation,”arXiv preprint arXiv:2509.21377, 2025

  12. [20]

    Beyond textual knowledge: Leverag- ing multimodal knowledge bases for enhancing vision-and-language navigation,

    D. Yang, Y . Yu, and L. Wang, “Beyond textual knowledge: Leverag- ing multimodal knowledge bases for enhancing vision-and-language navigation,”Information Processing & Management, vol. 63, no. 6, p. 104766, 2026

  13. [21]

    Advancing audio- visual navigation through multi-agent collaboration in 3d environ- ments,

    H. Zhang, Y . Yu, L. Wang, F. Sun, and W. Zheng, “Advancing audio- visual navigation through multi-agent collaboration in 3d environ- ments,” inInternational Conference on Neural Information Process- ing. Springer, 2025, pp. 502–516

  14. [22]

    Modality-invariant bidirec- tional temporal representation distillation network for missing multi- modal sentiment analysis,

    X. Wang, L. Wang, Y . Yu, and X. Jiao, “Modality-invariant bidirec- tional temporal representation distillation network for missing multi- modal sentiment analysis,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 202...

  15. [23]

    Vnet: A gan-based multi-tier discriminator network for speech synthesis vocoders,

    Y . Cao, Y . Li, L. Wang, and Y . Yu, “Vnet: A gan-based multi-tier discriminator network for speech synthesis vocoders,” in2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2024, pp. 4384–4389

  16. [24]

    Fsdenet: A frequency and spatial do- mains based detail enhancement network for remote sensing semantic segmentation,

    J. Fu, Y . Yu, and L. Wang, “Fsdenet: A frequency and spatial do- mains based detail enhancement network for remote sensing semantic segmentation,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2025

  17. [25]

    Bss-cffma: cross-domain feature fusion and multi-attention speech enhancement network based on self- supervised embedding,

    A. Mattursun, L. Wang, and Y . Yu, “Bss-cffma: cross-domain feature fusion and multi-attention speech enhancement network based on self- supervised embedding,” in2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE, 2024, pp. 3589–3594

  18. [26]

    Nonlinear regularization decoding method for speech recognition,

    J. Zhang, L. Wang, Y . Yu, and M. Xu, “Nonlinear regularization decoding method for speech recognition,”Sensors, vol. 24, no. 12, p. 3846, 2024

  19. [27]

    Iterative residual cross-attention mechanism: An integrated approach for audio-visual navigation tasks,

    H. Zhang, Y . Yu, L. Wang, F. Sun, and W. Zheng, “Iterative residual cross-attention mechanism: An integrated approach for audio-visual navigation tasks,”arXiv preprint arXiv:2509.25652, 2025

  20. [28]

    Recurrent neural networks,

    L. R. Medsker, L. Jainet al., “Recurrent neural networks,”Design and applications, vol. 5, no. 64-67, p. 2, 2001

  21. [29]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

  22. [30]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antigaet al., “Pytorch: An imperative style, high-performance deep learning library,”Advances in neural information processing systems, vol. 32, 2019

  23. [31]

    Introduction to convolutional neural networks,

    J. Wu, “Introduction to convolutional neural networks,”National Key Lab for Novel Software Technology. Nanjing University. China, vol. 5, no. 23, p. 495, 2017

  24. [32]

    The replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Vermaet al., “The replica dataset: A digital replica of indoor spaces,”arXiv preprint arXiv:1906.05797, 2019

  25. [33]

    Matterport3d: Learning from rgb-d data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niebner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3d: Learning from rgb-d data in indoor environments,” in7th IEEE International Conference on 3D Vision, 3DV 2017. Institute of Electrical and Electronics Engineers Inc...

  26. [34]

    On evaluation of embodied navigation agents,

    P. Anderson, A. Chang, D. S. Chaplot, A. Dosovitskiy, S. Gupta, V . Koltun, J. Kosecka, J. Malik, R. Mottaghi, M. Savvaet al., “On evaluation of embodied navigation agents,”arXiv preprint arXiv:1807.06757, 2018

  27. [2021]

    Virtual Event, Austria: OpenReview.net, May 3–7 2021

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.