pith. sign in

arxiv: 2605.19690 · v1 · pith:FU5627QHnew · submitted 2026-05-19 · 💻 cs.RO

D-CLING: Prior-Preserving Depth-Conditioned Fine-Tuning for Navigation Foundation Models

Shintaro Nakaoka , Takayuki Kanai , Kazuhito Tanaka This is my paper

Pith reviewed 2026-05-20 05:40 UTC · model grok-4.3

classification 💻 cs.RO
keywords navigation foundation modelsfine-tuningdepth conditioningprior preservationrobot navigationresidual pathwaysvisuomotor policycontinual learning
0
0 comments X

The pith

Attaching a zero-initialized residual copy of a pre-trained navigation backbone lets models learn in-domain geometry without eroding general priors.

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

Navigation foundation models trained on large datasets generalize well but lose robustness when fine-tuned on small in-domain sets, often resulting in more collisions and failed goals. The paper introduces D-CLING, which copies the pre-trained visual backbone and connects it through zero-initialized residual pathways to inject depth-based geometric cues. This structure lets the model acquire environment-specific obstacle and layout knowledge while the original pathways continue to supply broad behavioral competence. Real-world tests show the approach supports longer, safer navigation runs with less human help. Offline checks confirm action prediction accuracy stays the same or improves even on data outside the fine-tuning set.

Core claim

By freezing the original pre-trained weights and training only a parallel copy linked by zero-initialized residuals, the fine-tuning step adds depth-conditioned geometric adjustments that correct for new camera setups or environments while the preserved prior pathways keep the model from losing cross-domain navigation competence.

What carries the argument

Zero-initialized residual pathways that link a trainable duplicate of the pre-trained visual backbone to the frozen original network, enabling additive depth-conditioned learning.

If this is right

  • Long-horizon navigation runs become feasible with fewer obstacles hit and less need for human overrides.
  • Action prediction accuracy on held-out data stays stable or rises after the fine-tuning step.
  • The same structure supports adaptation to changed camera placements or new indoor layouts without full retraining.
  • Continual learning for navigation models becomes practical because prior knowledge is explicitly guarded during updates.

Where Pith is reading between the lines

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

  • The residual-copy design may transfer to other visuomotor foundation models that need domain-specific calibration without catastrophic forgetting.
  • Similar zero-init residual blocks could be tested on non-navigation tasks such as manipulation or locomotion to check whether the preservation effect generalizes.
  • Measuring how the residual pathways evolve during training might reveal the exact balance between geometric specialization and retained generality.
  • The method hints that explicit separation of prior and adaptation pathways could reduce the data needed for safe robot deployment in new sites.

Load-bearing premise

The copied backbone connected by zero-initialized residuals can absorb new geometric details without disrupting the behaviors already encoded in the pre-trained weights.

What would settle it

A head-to-head real-world trial in which the D-CLING model produces more collisions, more goal failures, or more human interventions than either the untouched pre-trained model or a conventionally fine-tuned version on the same novel setup.

Figures

Figures reproduced from arXiv: 2605.19690 by Kazuhito Tanaka, Shintaro Nakaoka, Takayuki Kanai.

Figure 1
Figure 1. Figure 1: Failure scenarios of the zero-shot NoMaD [6]. In real-robot navi￾gation, failures including unsafe clearance (left) and distance misestimation (right) were observed. In particular, the pre-training images [9, 10, 11, 12] exhibited strong distortion relative to the experimental condition, impairing geometric perception. Fine-tuned [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Less divergent trajectory generation after fine-tuning. Based on a shared start goal and an observation point, N = 20 trajectories sampled from each model are shown: Left: the zero-shot NoMaD (before fine￾tuning); Right: the fine-tuned model. Fine-tuning yielded markedly lower diversity, with a narrower spatial distribution and reduced heading variance, indicating a collapse of the pre-trained priors. attr… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed method: prior-preserving dense-depth conditioning, inspired by ControlNet [13]. Zero-initialized residual pathways inject per-pixel depth into intermediate features, preserving the pre-trained prior and improving geometry-consistent obstacle avoidance across cameras. LearnIng for General NaviGation Models), that leverages large-scale pre-training while efficiently adapting to novel… view at source ↗
Figure 4
Figure 4. Figure 4: D-CLING architectural design, based on NoMaD [6]. A frozen pre-trained RGB Branch (identical to the NoMaD architecture) maps an RGB history and goal image to intermediate features. In parallel, a trainable Depth Branch ingests RGB–D (with a 4→3 embedding) and injects zero-initialized residual features into the U-Net-based diffusion model; the two streams are fused by element-wise addition at each U-Net sta… view at source ↗
Figure 5
Figure 5. Figure 5: Representative frames of the proposed method from real-world experiments with a physical overlaid robot in an office environment: (i) Basic Obstacle—corridor traversal with visual avoidance of a single stationary box; (ii) Dynamic Corridor—after 10 m the robot must avoid an unmapped chair; and (iii) Long-range—a 50 m semicircular route across two junctions. label=(iii) 1) Basic obstacle avoidance (Basic Ob… view at source ↗
Figure 7
Figure 7. Figure 7: Representative real-world navigation examples of the RGB-D modality (Ours) in two scenarios. (i) Single Obstacle: corridor traversal with a single stationary chair. (ii) Multi Obstacle: avoidance of three obstacles in a zig-zag trajectory. TABLE III RGB VS. RGB-D CONDITIONING. SUCCESS RATE (SR) FOR TWO SCENARIOS. Modality (i) Single Obstacle SR(%)↑ (ii) Multi Obstacle SR(%)↑ RGB 60 10 RGB-D (Ours) 80 100 R… view at source ↗
Figure 6
Figure 6. Figure 6: Top-view of the predicted waypoint sequences. Waypoint predictions of length H + 1 = 8 generated from frame windows of length T +1 = 4 on a representative subset of the offline evaluation set. Each panel overlays D-CLING (blue) and the baselines of NoMaD (cyan), NoMaD￾FT (pink), and NoMaD-EF (gray) with the ground-truth path in green and start/goal markers. D-CLING most closely follows the ground truth and… view at source ↗
read the original abstract

Navigation Foundation Models (NFMs) trained on large cross-embodied datasets have demonstrated powerful generalizability in various scenarios. Adopting in-domain fine-tuning for an NFM efficiently calibrates the visuomotor policy, promising further improvement even in a novel scenario. However, the fine-tuned models still suffer from poor obstacle avoidance or fail to properly reach the provided goals. Furthermore, model updates using a small subset of data typically erode the pre-trained prior, compromising the pre-training generalization. Consequently, fine-tuning deteriorates the capability of the model for robust and accurate navigation. In this work, we present a novel fine-tuning method that leverages large-scale pre-training while efficiently learning in novel setups, such as environments or camera configurations. In particular, inspired by ControlNet, we fine-tune an NFM by attaching a trainable copy of the pre-trained backbone using zero-initialized residual pathways, thereby learning geometric cues. This design enables the model to efficiently acquire in-domain geometry while preserving pre-trained knowledge across various behaviors. Despite its simplicity, our comprehensive evaluation of real-world navigation suggests that our proposal effectively enables robust long-horizon navigation with minimal collisions and human intervention. Additionally, our offline analysis shows that the proposed method maintains or further improves action prediction capabilities beyond the fine-tuned dataset, providing a key insight into continual learning for general navigation. The project page: https://toyotafrc.github.io/DCLING-Proj/

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes D-CLING, a fine-tuning technique for Navigation Foundation Models (NFMs) that attaches a trainable copy of the pre-trained backbone via zero-initialized residual pathways, inspired by ControlNet. This design is intended to enable efficient acquisition of in-domain geometric cues (e.g., from depth conditioning) in novel environments or camera configurations while preserving pre-trained cross-behavior priors. The authors claim that the resulting models achieve robust real-world long-horizon navigation with minimal collisions and human intervention, and that offline analysis shows the method maintains or improves action-prediction performance beyond the fine-tuned dataset.

Significance. If the central claims are substantiated, the work would provide a simple architectural safeguard for continual learning in visuomotor navigation policies, addressing the common tension between domain adaptation and retention of large-scale pre-training generalization. The emphasis on depth-conditioned fine-tuning and the ControlNet-style residual attachment could be a practical contribution for deploying NFMs across diverse robotic setups.

major comments (2)
  1. [Abstract] Abstract: The claims of 'robust long-horizon navigation with minimal collisions and human intervention' and 'maintains or further improves action prediction capabilities beyond the fine-tuned dataset' are asserted without any quantitative metrics, baselines, ablation studies, or error analysis. This absence makes it impossible to evaluate whether the data support the central claims.
  2. [Method] Method section (ControlNet-inspired attachment): The key assumption that zero-initialized residual pathways preserve pre-trained generalization on out-of-domain behaviors lacks supporting controls. No before/after comparisons on held-out pre-training distributions, no ablation removing the zero-init or residual structure, and no direct contrast to standard fine-tuning (which the paper states erodes priors) are provided. Without these, it remains possible that any robustness derives from the fine-tuning data distribution rather than the architectural choice.
minor comments (2)
  1. [Abstract] The project page URL is given but no statement on code or model release appears in the manuscript; adding this would aid reproducibility.
  2. [Method] Notation for the residual pathways and depth conditioning could be clarified with a diagram or explicit equations in the method description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive feedback on our manuscript. We address each major comment below and outline revisions that will strengthen the presentation of our results and the supporting evidence for the architectural design.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claims of 'robust long-horizon navigation with minimal collisions and human intervention' and 'maintains or further improves action prediction capabilities beyond the fine-tuned dataset' are asserted without any quantitative metrics, baselines, ablation studies, or error analysis. This absence makes it impossible to evaluate whether the data support the central claims.

    Authors: We agree that the abstract would benefit from explicit quantitative support. The full manuscript reports real-world navigation results across multiple environments and camera setups, including success rates, collision counts, and human intervention frequency relative to the pre-trained baseline and standard fine-tuning. Offline action-prediction accuracy is also evaluated on held-out sequences. We will revise the abstract to include the key numerical outcomes (e.g., success rate, average collisions per episode) so that the central claims are directly substantiated by the reported metrics. revision: yes

  2. Referee: [Method] Method section (ControlNet-inspired attachment): The key assumption that zero-initialized residual pathways preserve pre-trained generalization on out-of-domain behaviors lacks supporting controls. No before/after comparisons on held-out pre-training distributions, no ablation removing the zero-init or residual structure, and no direct contrast to standard fine-tuning (which the paper states erodes priors) are provided. Without these, it remains possible that any robustness derives from the fine-tuning data distribution rather than the architectural choice.

    Authors: This is a fair observation. While the current offline analysis shows that D-CLING maintains or improves action prediction on sequences outside the fine-tuning distribution, we did not include explicit before/after evaluations on the original pre-training corpus or systematic ablations of the zero-initialization and residual structure. We will add these controls to the revised manuscript: (1) action-prediction accuracy on held-out pre-training data before and after D-CLING fine-tuning, (2) an ablation removing zero-initialization, and (3) a direct comparison against standard fine-tuning without the residual pathways. These additions will isolate the contribution of the architectural choice. revision: yes

Circularity Check

0 steps flagged

No circularity; method is a design choice grounded in external ControlNet inspiration

full rationale

The paper describes a fine-tuning architecture that attaches a trainable copy of the pre-trained backbone via zero-initialized residual pathways, explicitly inspired by ControlNet. This is presented as an engineering choice to acquire in-domain geometric cues while preserving cross-behavior priors from large-scale pre-training. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the provided text. Claims rest on real-world navigation trials and offline action-prediction analysis rather than reducing by construction to inputs defined within the paper itself. The approach is therefore self-contained against external benchmarks and pre-trained models.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the pre-trained NFM encodes reusable navigation priors that can be protected during targeted geometric adaptation; no explicit free parameters or invented entities are identifiable from the abstract alone.

axioms (1)
  • domain assumption Pre-trained navigation foundation models contain generalizable visuomotor knowledge that remains useful after targeted fine-tuning.
    Invoked to justify why preserving the prior is valuable and why the residual design succeeds.

pith-pipeline@v0.9.0 · 5796 in / 1207 out tokens · 41922 ms · 2026-05-20T05:40:55.622192+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

46 extracted references · 46 canonical work pages

  1. [1]

    Target-driven visual navigation in indoor scenes using deep reinforcement learning

    Y . Zhu, R. Mottaghi, E. Kolve, J. J. Lim, A. Gupta, L. Fei-Fei, and A. Farhadi, “Target-driven visual navigation in indoor scenes using deep reinforcement learning.” InProc. of the ICRA, 2017, pp. 3357–3364

    work page 2017

  2. [2]

    Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames,

    E. Wijmans, A. Kadian, A. Morcos, S. Lee, I. Essa, D. Parikh, M. Savva, and D. Batra, “Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames,”arXiv, 2019

    work page 2019

  3. [3]

    Socially compliant navigation through raw depth inputs with generative adversarial imitation learning

    L. Tai, J. Zhang, M. Liu, and W. Burgard, “Socially compliant navigation through raw depth inputs with generative adversarial imitation learning.” InProc. of the ICRA, 2018, pp. 1111–1117

    work page 2018

  4. [4]

    Habitat-web: Learning embodied object-search strategies from human demonstra- tions at scale

    R. Ramrakhya, E. Undersander, D. Batra, and A. Das, “Habitat-web: Learning embodied object-search strategies from human demonstra- tions at scale.” InProc. of the CVPR, 2022, pp. 5173–5183

    work page 2022

  5. [5]

    Vint: A foundation model for visual navigation,

    D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine, “Vint: A foundation model for visual navigation,”arXiv, 2023

    work page 2023

  6. [6]

    Nomad: Goal masked diffusion policies for navigation and exploration

    A. Sridhar, D. Shah, C. Glossop, and S. Levine, “Nomad: Goal masked diffusion policies for navigation and exploration.” InProc. of the ICRA, 2024, pp. 63–70

    work page 2024

  7. [7]

    Data scaling for navigation in unknown envi- ronments,

    L. Suomela, N. Takahata, S. K. Arachchige, H. Edelman, and J.-K. K ¨am¨ar¨ainen, “Data scaling for navigation in unknown envi- ronments,”arXiv, 2026

    work page 2026

  8. [8]

    Pig-nav: Key insights for pretrained image goal navigation models,

    J. Wan, C. Zhou, J. Liu, X. Huang, X. Chen, X. Yi, Q. Yang, B. Zhu, X.-Q. Cai, L. Liu, et al., “Pig-nav: Key insights for pretrained image goal navigation models,”arXiv, 2025

    work page 2025

  9. [9]

    Sacson: Scalable autonomous control for social navigation,

    N. Hirose, D. Shah, A. Sridhar, and S. Levine, “Sacson: Scalable autonomous control for social navigation,”IEEE RA-L, vol. 9, no. 1, pp. 49–56, 2023

    work page 2023

  10. [10]

    Rapid exploration for open-world navigation with latent goal models,

    D. Shah, B. Eysenbach, G. Kahn, N. Rhinehart, and S. Levine, “Rapid exploration for open-world navigation with latent goal models,”arXiv, 2021

    work page 2021

  11. [11]

    Deep visual mpc-policy learning for navigation,

    N. Hirose, F. Xia, R. Mart ´ın-Mart´ın, A. Sadeghian, and S. Savarese, “Deep visual mpc-policy learning for navigation,”IEEE RA-L, vol. 4, no. 4, pp. 3184–3191, 2019

    work page 2019

  12. [12]

    Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation,

    H. Karnan, A. Nair, X. Xiao, G. Warnell, S. Pirk, A. Toshev, J. Hart, J. Biswas, and P. Stone, “Socially compliant navigation dataset (scand): A large-scale dataset of demonstrations for social navigation,”IEEE RA-L, vol. 7, no. 4, pp. 11 807–11 814, 2022

    work page 2022

  13. [13]

    Adding conditional control to text-to-image diffusion models

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models.” InProc. of the CVPR, 2023, pp. 3836–3847

    work page 2023

  14. [14]

    Flownav: Combining flow matching and depth priors for efficient navigation,

    S. Gode, A. Nayak, D. N. Oliveira, M. Krawez, C. Schmid, and W. Burgard, “Flownav: Combining flow matching and depth priors for efficient navigation,”arXiv, 2024

    work page 2024

  15. [15]

    Mapping, localization and path planning for image-based naviga- tion using visual features and map

    J. Thoma, D. P. Paudel, A. Chhatkuli, T. Probst, and L. V . Gool, “Mapping, localization and path planning for image-based naviga- tion using visual features and map.” InProc. of the CVPR, 2019, pp. 7383–7391

    work page 2019

  16. [16]

    Orb-slam: A versatile and accurate monocular slam system,

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos, “Orb-slam: A versatile and accurate monocular slam system,”IEEE transactions on robotics, vol. 31, no. 5, pp. 1147–1163, 2015

    work page 2015

  17. [17]

    WayFAST: Navigation With Predictive Traversability in the Field,

    M. V . Gasparino, A. N. Sivakumar, Y . Liu, A. E. B. Velasquez, V . A. H. Higuti, J. Rogers, H. Tran, and G. Chowdhary, “WayFAST: Navigation With Predictive Traversability in the Field,”IEEE RA-L, vol. 7, no. 4, pp. 10 651–10 658, 2022

    work page 2022

  18. [18]

    Learning to explore using active neural slam

    D. S. Chaplot, D. Gandhi, S. Gupta, A. Gupta, and R. Salakhutdinov, “Learning to explore using active neural slam.” InProc. of the ICLR, 2020

    work page 2020

  19. [19]

    Lifelong autonomous improvement of navigation foundation models in the wild

    K. Stachowicz, L. Ignatova, and S. Levine, “Lifelong autonomous improvement of navigation foundation models in the wild.” InProc. of the CoRL, 2024

    work page 2024

  20. [20]

    Learning hierarchical relationships for object-goal navigation

    Y . Qiu, A. Pal, and H. I. Christensen, “Learning hierarchical relationships for object-goal navigation.” InProc. of the CoRL, 2020

    work page 2020

  21. [21]

    Creste: Scalable mapless navigation with internet scale priors and counterfactual guidance

    A. Zhang, H. Sikchi, A. Zhang, and J. Biswas, “Creste: Scalable mapless navigation with internet scale priors and counterfactual guidance.” InProc. of the RSS, 2025

    work page 2025

  22. [22]

    Citywalker: Learning embodied urban navigation from web-scale videos

    X. Liu, J. Li, Y . Jiang, N. Sujay, Z. Yang, J. Zhang, J. Abanes, J. Zhang, and C. Feng, “Citywalker: Learning embodied urban navigation from web-scale videos.” InProc. of the CVPR, 2025, pp. 6875–6885

    work page 2025

  23. [23]

    Image-goal navigation using refined feature guidance and scene graph enhancement,

    Z. Feng, X. Chen, C. Shi, L. Luo, Z. Chen, Y .-H. Liu, and H. Lu, “Image-goal navigation using refined feature guidance and scene graph enhancement,”arXiv, 2025

    work page 2025

  24. [24]

    Vision- and-language navigation: Interpreting visually-grounded navigation instructions in real environments

    P. Anderson, Q. Wu, D. Teney, J. Bruce, M. Johnson, N. S¨underhauf, I. Reid, S. Gould, and A. Van Den Hengel, “Vision- and-language navigation: Interpreting visually-grounded navigation instructions in real environments.” InProc. of the CVPR, 2018, pp. 3674–3683

    work page 2018

  25. [25]

    A new path: Scaling vision- and-language navigation with synthetic instructions and imitation learning

    A. Kamath, P. Anderson, S. Wang, J. Y . Koh, A. Ku, A. Waters, Y . Yang, J. Baldridge, and Z. Parekh, “A new path: Scaling vision- and-language navigation with synthetic instructions and imitation learning.” InProc. of the CVPR, 2023, pp. 10 813–10 823

    work page 2023

  26. [26]

    Lelan: Learning a language-conditioned navigation policy from in- the-wild videos,

    N. Hirose, C. Glossop, A. Sridhar, D. Shah, O. Mees, and S. Levine, “Lelan: Learning a language-conditioned navigation policy from in- the-wild videos,”arXiv, 2024

    work page 2024

  27. [27]

    Vlm-social-nav: Socially aware robot navigation through scoring using vision-language models,

    D. Song, J. Liang, A. Payandeh, A. H. Raj, X. Xiao, and D. Manocha, “Vlm-social-nav: Socially aware robot navigation through scoring using vision-language models,”IEEE RA-L, 2024

    work page 2024

  28. [28]

    Diagnosing the environment bias in vision-and-language navigation

    Y . Zhang, H. Tan, and M. Bansal, “Diagnosing the environment bias in vision-and-language navigation.” InProc. of the IJCAI, 2020, pp. 890–897

    work page 2020

  29. [29]

    Monocular camera- based point-goal navigation by learning depth channel and cross- modality pyramid fusion

    X. Y . Tianqi Tang Heming Du and Y . Yang, “Monocular camera- based point-goal navigation by learning depth channel and cross- modality pyramid fusion.” InProc. of the AAAI, 2022, 36(5), 5422–5430

    work page 2022

  30. [30]

    Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance,

    W. Cai, J. Peng, Y . Yang, Y . Zhang, M. Wei, H. Wang, Y . Chen, T. Wang, and J. Pang, “Navdp: Learning sim-to-real navigation diffusion policy with privileged information guidance,”arXiv, 2025

    work page 2025

  31. [31]

    X-nav: Learning end-to-end cross-embodiment navigation for mobile robots,

    H. Wang, A. H. Tan, A. Fung, and G. Nejat, “X-nav: Learning end-to-end cross-embodiment navigation for mobile robots,”arXiv, 2025

    work page 2025

  32. [32]

    A careful examination of large behavior models for multitask dexterous manipulation,

    T. L. Team, “A careful examination of large behavior models for multitask dexterous manipulation,”arXiv, 2025

    work page 2025

  33. [33]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al., “Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0.” InProc. of the ICRA, IEEE, 2024, pp. 6892–6903

    work page 2024

  34. [34]

    π0.5: A vision-language- action model with open-world generalization,

    K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, et al., “π0.5: A vision-language- action model with open-world generalization,”arXiv, 2025

    work page 2025

  35. [35]

    Gr00t n1: An open foundation model for generalist humanoid robots,

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al., “Gr00t n1: An open foundation model for generalist humanoid robots,”arXiv, 2025

    work page 2025

  36. [36]

    Tail: Task-specific adapters for imitation learning with large pretrained models

    Z. Liu, J. Zhang, K. Asadi, Y . Liu, D. Zhao, S. Sabach, and R. Fakoor, “Tail: Task-specific adapters for imitation learning with large pretrained models.” InProc. of the ICLR, 2024

    work page 2024

  37. [37]

    Learning generalizable manipulation policy with adapter-based parameter fine-tuning

    K. Lu, K. T. Ly, W. Hebberd, K. Zhou, I. Havoutis, and A. Markham, “Learning generalizable manipulation policy with adapter-based parameter fine-tuning.” InProc. of the IROS, 2024

    work page 2024

  38. [38]

    Lossless adaptation of pretrained vision models for robotic manipulation,

    M. Sharma, C. Fantacci, Y . Zhou, S. Koppula, N. Heess, J. Scholz, and Y . Aytar, “Lossless adaptation of pretrained vision models for robotic manipulation,”arXiv, 2023

    work page 2023

  39. [39]

    Robocat: A self- improving generalist agent for robotic manipulation,

    K. Bousmalis, G. Vezzani, D. Rao, C. Devin, A. X. Lee, M. Bauz ´a, T. Davchev, Y . Zhou, A. Gupta, A. Raju, et al., “Robocat: A self- improving generalist agent for robotic manipulation,”arXiv, 2023

    work page 2023

  40. [40]

    Enhancing safety of foundation models for visual navigation through collision avoidance via repulsive estimation,

    J. Kim, J. Sim, W. Kim, K. Sycara, and C. Nam, “Enhancing safety of foundation models for visual navigation through collision avoidance via repulsive estimation,”arXiv, 2025

    work page 2025

  41. [41]

    Gsplatvnm: Point-of-view synthesis for visual navigation models using gaussian splatting,

    K. Honda, T. Ishita, Y . Yoshimura, and R. Yonetani, “Gsplatvnm: Point-of-view synthesis for visual navigation models using gaussian splatting,”arXiv, 2025

    work page 2025

  42. [42]

    Development of Human Support Robot as the research platform of a domestic mobile manipulator,

    T. Yamamoto, K. Terada, A. Ochiai, F. Saito, Y . Asahara, and K. Murase, “Development of Human Support Robot as the research platform of a domestic mobile manipulator,”ROBOMECH J., vol. 6, no. 1, p. 4, 2019

    work page 2019

  43. [43]

    A learned stereo depth system for robotic manipulation in homes,

    K. Shankar, M. Tjersland, J. Ma, K. Stone, and M. Bajracharya, “A learned stereo depth system for robotic manipulation in homes,” IEEE RA-L, vol. 7, no. 2, pp. 2305–2312, 2022

    work page 2022

  44. [44]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,”arXiv, 2017

    work page 2017

  45. [45]

    Depth anything v2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth anything v2,”NeurIPS, vol. 37, pp. 21 875–21 911, 2024

    work page 2024

  46. [46]

    General evaluation for instruction conditioned navigation using dynamic time warping,

    G. Ilharco, V . Jain, A. Ku, E. Ie, and J. Baldridge, “General evaluation for instruction conditioned navigation using dynamic time warping,”arXiv, 2019

    work page 2019