arxiv: 2604.24391 · v1 · submitted 2026-04-27 · 💻 cs.RO

Recognition: unknown

FreqCache: Accelerating Embodied VLN Models with Adaptive Frequency-Guided Token Caching

Zihao Zheng , Xingyue Zhou , Zhihao Mao , Songyu Sun , Lingyue Zhang , Yulong Ao , Yupu Feng , Qiongqiong Zhang

show 2 more authors

Yonghua Lin Xiang Chen

Authors on Pith no claims yet

Pith reviewed 2026-05-08 03:06 UTC · model grok-4.3

classification 💻 cs.RO

keywords vision-language navigationtoken cachingfrequency domainmodel accelerationembodied AIadaptive cachingcomputational efficiency

0 comments

The pith

Frequency domain analysis lets token caching accelerate VLN models by handling viewpoint shifts and temporal changes that break visual methods.

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

The paper establishes that existing token caching for vision-language navigation fails under viewpoint migration, edge neglect, and missing temporal adjustment, but these problems remain consistent and measurable in the frequency domain. This consistency supports a new framework that sets up caches, refreshes them, and tunes budgets automatically using frequency properties. A sympathetic reader would care because VLN models navigate well yet run slowly on embodied hardware, so improved caching could cut compute costs without retraining or accuracy loss. If the analysis holds, frequency-domain selection becomes a practical route to faster inference in dynamic visual environments.

Core claim

Detailed analyses reveal that the impacts of viewpoint migration, edge-information neglect, and lack of temporal adjustability exhibit invariance and analyzability in the frequency domain. Based on these, we propose a frequency-guided token caching framework, called FreqCache. Utilizing the inherent properties of the frequency domain, FreqCache achieves optimal token cache establishment, refreshment, and adaptive adjustment. Experiments show that FreqCache achieves 1.59x speedup with ignorable overhead.

What carries the argument

FreqCache, the frequency-guided token caching framework that exploits invariance of challenge impacts in the frequency domain to decide which tokens to cache, when to refresh them, and how to adjust budgets.

Load-bearing premise

The impacts of viewpoint migration, edge-information neglect, and lack of temporal adjustability exhibit invariance and analyzability in the frequency domain.

What would settle it

Apply FreqCache to VLN episodes containing rapid viewpoint changes and measure whether navigation success rate drops or speedup falls below 1.2x compared with uncached baselines.

Figures

Figures reproduced from arXiv: 2604.24391 by Lingyue Zhang, Qiongqiong Zhang, Songyu Sun, Xiang Chen, Xingyue Zhou, Yonghua Lin, Yulong Ao, Yupu Feng, Zhihao Mao, Zihao Zheng.

**Figure 1.** Figure 1: VLN Token Caching Challenges and Comparison between Visual Domain Methods and the Proposed view at source ↗

**Figure 2.** Figure 2: Domain Discrepancy in Viewpoint Migration view at source ↗

**Figure 3.** Figure 3: Domain Discrepancy in Edge Identification view at source ↗

**Figure 5.** Figure 5: Details of the proposed FreqCache Framework during navigation, we introduce spectral entropy Ψ 𝑡 as an intrinsic proxy for scene complexity and propose the Temporal Adaptive Cache Budget Determination module. As analyzed in Sec. 3, spectral entropy’s changes can correspond to scenario temporal variations. In structurally simple hallway stages, energy concentrates in low frequencies (low spectral entropy),… view at source ↗

**Figure 6.** Figure 6: System Implementation of FreqCache Framework requirement for real-time navigation, we select GPU as the primary computing device. We use almost 5000 lines of code to implement FreqCache, based on Pytorch library and CUDA architecture. For the other two patterns, we also implement them on GPUs rather than CPUs to avoid the high CPU-GPU data transfer overhead. Given the computational characteristics, tensor… view at source ↗

**Figure 7.** Figure 7: Token Reuse Visualization on an Episode from R2R-CE view at source ↗

**Figure 8.** Figure 8: Discussion of Hyperparameters in FreqCache presented in view at source ↗

read the original abstract

Vision-Language-Navigation (VLN) models exhibit excellent navigation accuracy but incur high computational overhead. Token caching has emerged as a promising training-free strategy to reduce this cost by reusing token computation results; however, existing token caching approaches rely on visual domain methods for cacheable token selection, leading to challenges when adapted to VLN models. 1) Visual domain methods become invalid when there is viewpoint migration. 2) Visual domain methods neglect critical edge information without the aid of additional algorithms. 3) Visual domain methods overlook the temporal variation of scenarios and lack adjustability in cache budgets. In this paper, we develop detailed analyses and find that the impacts of these challenges exhibit invariance and analyzability in the frequency domain. Based on these, we propose a frequency-guided token caching framework, called FreqCache. Utilizing the inherent properties of the frequency domain, FreqCache achieves optimal token cache establishment, refreshment, and adaptive adjustment. Experiments show that FreqCache achieves 1.59x speedup with ignorable overhead, showing the effect of integrating frequency domain methods in VLN token caching.

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.

Desk Editor's Note private letter to a colleague

FreqCache applies frequency-domain analysis to fix specific token-caching failures in VLN, but the invariance claims need tighter validation to support the 1.59x speedup.

read the letter

FreqCache moves token caching for vision-language navigation into the frequency domain to handle viewpoint migration, neglected edges, and temporal changes better than visual-domain methods, with a claimed 1.59x speedup and negligible overhead. The authors identify three concrete problems that arise when standard caching is ported to VLN and argue that frequency properties give the needed stability for cache establishment, refresh, and adaptive budgets. That framing is the clearest new element. They also do a reasonable job explaining why visual approaches break in navigation settings and why frequency might supply invariance without extra algorithms. The practical efficiency angle is worth attention if the numbers hold across models. The main soft spot is the central assumption. The paper states that detailed analyses confirm the impacts of the three challenges are invariant and analyzable in frequency, yet the provided details leave the actual transforms, stability metrics under viewpoint shifts, and recovery of edge information via high-frequency bands unclear. Without those steps shown explicitly, it is hard to tell whether frequency guidance is doing the real work or whether the gains come from a different heuristic. Experiments would also benefit from fuller baselines, ablations on each component, and dataset specifics to make the speedup convincing. The citation pattern is standard for the area. This paper is for researchers working on efficient inference for embodied VLN agents. A reader who needs training-free speedups on navigation models would get usable ideas from the adaptive framework. It deserves a serious referee because the computational bottleneck is real and the frequency angle is distinct enough to check in detail. I would send it to review.

Referee Report

1 major / 2 minor

Summary. The paper proposes FreqCache, a frequency-guided token caching framework for Vision-Language Navigation (VLN) models. It identifies three challenges with existing visual-domain token caching methods (invalidity under viewpoint migration, neglect of edge information, and lack of temporal adjustability in cache budgets) and claims that detailed analyses show these impacts exhibit invariance and analyzability in the frequency domain. This enables optimal token cache establishment, refreshment, and adaptive adjustment, yielding a reported 1.59x speedup with negligible overhead.

Significance. If the frequency-domain invariance holds with explicit validation and the speedup is shown to be reproducible against standard baselines with controls for overhead, this could offer a training-free acceleration technique for embodied VLN models by introducing frequency-domain properties as a basis for caching decisions. The approach's potential to address viewpoint and temporal issues in a principled way would be a useful contribution to efficient inference in robotics and navigation tasks.

major comments (1)

[Abstract] Abstract: The central claim that 'the impacts of these challenges exhibit invariance and analyzability in the frequency domain' (enabling FreqCache's cache logic) is load-bearing for both the method and the 1.59x speedup result, yet the text provides no frequency transforms, equations (e.g., Fourier analysis or stability metrics across views), or quantitative invariance tests. Without this derivation or validation, the frequency guidance reduces to an unverified heuristic and the speedup claim lacks a demonstrated causal link to the frequency properties.

minor comments (2)

[Abstract] Abstract: The speedup is stated as '1.59x' without reference to specific datasets, baselines, number of runs, or error bars; this should be expanded in the experiments section with a clear protocol for reproducibility.
[Abstract] Abstract: 'Ignorable overhead' is vague and should be replaced with a quantified measurement (e.g., percentage increase in latency or memory) and comparison to the baseline.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the thorough review and valuable feedback on our work. We agree that the frequency-domain invariance claim is central and requires more explicit mathematical support and validation to strengthen the manuscript. We address the comment below and will incorporate revisions to provide the requested derivations and tests.

read point-by-point responses

Referee: [Abstract] Abstract: The central claim that 'the impacts of these challenges exhibit invariance and analyzability in the frequency domain' (enabling FreqCache's cache logic) is load-bearing for both the method and the 1.59x speedup result, yet the text provides no frequency transforms, equations (e.g., Fourier analysis or stability metrics across views), or quantitative invariance tests. Without this derivation or validation, the frequency guidance reduces to an unverified heuristic and the speedup claim lacks a demonstrated causal link to the frequency properties.

Authors: We appreciate this observation and agree that the presentation of the frequency-domain analysis can be strengthened for clarity and rigor. While the manuscript describes the analyses leading to the invariance insight, we acknowledge that explicit Fourier transforms, stability equations, and quantitative cross-view metrics are not sufficiently detailed in the current version. In the revised manuscript, we will add a dedicated subsection (likely in Section 3) that includes: (1) the Fourier transform formulations used to analyze token impacts, (2) mathematical derivations showing invariance of frequency components under viewpoint migration (e.g., phase and magnitude stability metrics), and (3) quantitative experiments with tables/figures reporting invariance scores across navigation sequences. We will also add an ablation linking these properties directly to cache selection/refresh decisions and the measured 1.59x speedup. This will establish the causal connection and move beyond any heuristic interpretation. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; derivation is self-contained.

full rationale

The paper identifies three visual-domain limitations in token caching for VLN, then states that frequency-domain analyses reveal invariance properties for these challenges, which directly motivate the design of FreqCache for cache establishment, refreshment, and adaptive adjustment. The 1.59x speedup is reported as an experimental outcome. No quoted equations or steps reduce the claimed frequency invariance to a fitted parameter, self-citation chain, or definitional tautology. The analyses are presented as independent observations feeding the method, with empirical validation outside the derivation itself. This satisfies the default expectation of non-circularity for most papers.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on one domain assumption about frequency-domain invariance; no free parameters or new entities are explicitly introduced in the abstract.

axioms (1)

domain assumption The impacts of viewpoint migration, edge-information neglect, and lack of temporal adjustability exhibit invariance and analyzability in the frequency domain.
This invariance is the explicit basis given in the abstract for building the frequency-guided caching rules.

pith-pipeline@v0.9.0 · 5524 in / 1225 out tokens · 73561 ms · 2026-05-08T03:06:05.162185+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

45 extracted references · 21 canonical work pages · 8 internal anchors

[1]

Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünder- hauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. 2018. Vision-and- language navigation: Interpreting visually-grounded navigation instructions in real environments. InProceedings of the IEEE conference on computer vision and pattern recognition. 3674–3683

2018
[2]

S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y. Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y. Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J. Lin. 2025. qwen25vl. arXiv:2502.13923 [cs.CV]

work page internal anchor Pith review arXiv 2025
[3]

Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. 2022. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817(2022)

work page internal anchor Pith review arXiv 2022
[4]

Jiaqi Chen, Bingqian Lin, Ran Xu, Zhenhua Chai, Xiaodan Liang, and Kwan-Yee Wong. 2024. Mapgpt: Map-guided prompting with adaptive path planning for vision-and-language navigation. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 9796–9810

2024
[5]

Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. 2024. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InEuropean Conference on Computer Vision. Springer, 19–35

2024
[6]

Shizhe Chen, Pierre-Louis Guhur, Cordelia Schmid, and Ivan Laptev. 2021. History aware multimodal transformer for vision-and-language navigation.Advances in neural information processing systems34 (2021), 5834–5847

2021
[7]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

work page internal anchor Pith review arXiv 2020
[8]

2009.Digital image processing

Rafael C Gonzalez. 2009.Digital image processing. Pearson education india

2009
[9]

Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. 2022. Vision- and-language navigation: A survey of tasks, methods, and future directions. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7606–7623

2022
[10]

Jialuo He and Huangxun Chen. 2026. Energy-Driven Adaptive Visual Token Pruning for Efficient Vision-Language Models.arXiv preprint arXiv:2603.05950 (2026)

work page arXiv 2026
[11]

Kai Huang, Hao Zou, Bochen Wang, Ye Xi, Zhen Xie, and Hao Wang. 2025. AirCache: Activating Inter-modal Relevancy KV Cache Compression for Effi- cient Large Vision-Language Model Inference. InProceedings of the IEEE/CVF International Conference on Computer Vision. 23958–23967

2025
[12]

Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee
[13]

InEuropean Conference on Computer Vision (ECCV)

Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments. InEuropean Conference on Computer Vision (ECCV)
[14]

Yanwei Li, Chengyao Wang, and Jiaya Jia. 2024. Llama-vid: An image is worth 2 tokens in large language models. InEuropean Conference on Computer Vision. Springer, 323–340

2024
[15]

Rui Liu, Xiaohan Wang, Wenguan Wang, and Yi Yang. 2023. Bird’s-Eye-View Scene Graph for Vision-Language Navigation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

2023
[16]

Yuxing Long, Wenzhe Cai, Hongcheng Wang, Guanqi Zhan, and Hao Dong. 2024. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment.arXiv preprint arXiv:2406.04882(2024)

work page arXiv 2024
[17]

1999.A wavelet tour of signal processing

Stéphane Mallat. 1999.A wavelet tour of signal processing. Elsevier

1999
[18]

Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. 2025. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747 (2025)

work page internal anchor Pith review arXiv 2025
[19]

Wenda Qin, Andrea Burns, Bryan A Plummer, and Margrit Betke. 2025. Walk and Read Less: Improving the Efficiency of Vision-and-Language Navigation via Tuning-Free Multimodal Token Pruning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 23567–23581

2025
[20]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al
[21]

In International conference on machine learning

Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 8748–8763
[22]

B Srinivasa Reddy and Biswanath N Chatterji. 1996. An FFT-based technique for translation, rotation, and scale-invariant image registration.IEEE transactions on image processing5, 8 (1996), 1266–1271

1996
[23]

Dhruv Shah, Błażej Osiński, Sergey Levine, et al. 2023. Lm-nav: Robotic naviga- tion with large pre-trained models of language, vision, and action. InConference on robot learning. pmlr, 492–504

2023
[24]

Kele Shao, TAO Keda, Kejia Zhang, Sicheng Feng, Mu Cai, Yuzhang Shang, Haoxuan You, Can Qin, Yang Sui, and Huan Wang. [n. d.]. A Survey of Token Compression for Efficient Multimodal Large Language Models.Transactions on Machine Learning Research([n. d.])
[25]

InternNav Team. 2025. InternVLA-N1: An Open Dual-System Navigation Foun- dation Model with Learned Latent Plans

2025
[26]

Dezhan Tu, Danylo Vashchilenko, Yuzhe Lu, and Panpan Xu. 2024. VL-cache: Sparsity and modality-aware KV cache compression for vision-language model inference acceleration.arXiv preprint arXiv:2410.23317(2024)

work page arXiv 2024
[27]

Huanyu Wang, Jushi Kai, Haoli Bai, Lu Hou, Bo Jiang, Ziwei He, and Zhouhan Lin. 2025. Fourier-vlm: Compressing vision tokens in the frequency domain for large vision-language models.arXiv preprint arXiv:2508.06038(2025)

work page arXiv 2025
[28]

Meng Wei, Chenyang Wan, Xiqian Yu, Tai Wang, Yuqiang Yang, Xiaohan Mao, Chenming Zhu, Wenzhe Cai, Hanqing Wang, Yilun Chen, et al. 2025. Streamvln: Streaming vision-and-language navigation via slowfast context modeling.arXiv preprint arXiv:2507.05240(2025)

work page arXiv 2025
[29]

Zhiqiang Xie, Ziyi Xu, Mark Zhao, Yuwei An, Vikram Sharma Mailthody, Scott Mahlke, Michael Garland, and Christos Kozyrakis. 2025. Strata: Hierarchi- cal context caching for long context language model serving.arXiv preprint arXiv:2508.18572(2025)

work page arXiv 2025
[32]

Siyu Xu, Yunke Wang, Chenghao Xia, Dihao Zhu, Tao Huang, and Chang Xu
[33]

VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching.arXiv preprint arXiv:2502.02175(2025)

work page arXiv 2025
[34]

Yaoxin Yang, Peng Ye, Xudong Tan, Chongjun Tu, Maosen Zhao, Jia Hao, and Tao Chen. 2025. Revisiting Multimodal KV Cache Compression: A Frequency- Domain-Guided Outlier-KV-Aware Approach.arXiv preprint arXiv:2511.16786 (2025)

work page arXiv 2025
[35]

Shuhao Ye, Sitong Mao, Yuxiang Cui, Xuan Yu, Shichao Zhai, Wen Chen, Shunbo Zhou, Rong Xiong, and Yue Wang. 2025. ETP-R1: Evolving Topological Planning with Reinforcement Fine-tuning for Vision-Language Navigation in Continuous Environments.arXiv preprint arXiv:2512.20940(2025)

work page arXiv 2025
[36]

Jiazhao Zhang, Kunyu Wang, Shaoan Wang, Minghan Li, Haoran Liu, Songlin Wei, Zhongyuan Wang, Zhizheng Zhang, and He Wang. 2024. Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks.arXiv preprint arXiv:2412.06224(2024)

work page arXiv 2024
[37]

Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and He Wang. 2024. Navid: Video-based vlm plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852 (2024)

work page arXiv 2024
[38]

Duo Zheng, Shijia Huang, Yanyang Li, and Liwei Wang. 2025. Efficient- VLN: A Training-Efficient Vision-Language Navigation Model.arXiv preprint arXiv:2512.10310(2025)

work page arXiv 2025
[39]

Duo Zheng, Shijia Huang, Lin Zhao, Yiwu Zhong, and Liwei Wang. 2024. To- wards learning a generalist model for embodied navigation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13624–13634

2024
[40]

Zihao Zheng, Hangyu Cao, Jiayu Chen, Sicheng Tian, Chenyue Li, Maoliang Li, Xinhao Sun, Guojie Luo, and Xiang Chen. 2026. RoboECC: Multi-Factor- Aware Edge-Cloud Collaborative Deployment for VLA Models.arXiv preprint arXiv:2603.20711(2026)

work page internal anchor Pith review Pith/arXiv arXiv 2026
[41]

Zihao Zheng, Zhihao Mao, Maoliang Li, Jiayu Chen, Xinhao Sun, Zhaobo Zhang, Donggang Cao, Hong Mei, and Xiang Chen. 2026. Kerv: Kinematic-rectified speculative decoding for embodied vla models.arXiv preprint arXiv:2603.01581 (2026)

work page internal anchor Pith review Pith/arXiv arXiv 2026
[42]

Zihao Zheng, Zhihao Mao, Sicheng Tian, Maoliang Li, Jiayu Chen, Xinhao Sun, Zhaobo Zhang, Xuanzhe Liu, Donggang Cao, Hong Mei, et al . 2026. HeiSD: Hybrid Speculative Decoding for Embodied Vision-Language-Action Models with Kinematic Awareness.arXiv preprint arXiv:2603.17573(2026)

work page internal anchor Pith review Pith/arXiv arXiv 2026
[43]

Zihao Zheng, Zhihao Mao, Xingyue Zhou, Jiayu Chen, Maoliang Li, Xinhao Sun, Hailong Zou, Zhaobo Zhang, Xuanzhe Liu, Donggang Cao, et al . 2026. VLN- Cache: Enabling Token Caching for VLN Models with Visual/Semantic Dynamics Awareness.arXiv preprint arXiv:2603.07080(2026)

work page internal anchor Pith review Pith/arXiv arXiv 2026
[44]

Zihao Zheng, Sicheng Tian, Hangyu Cao, Chenyue Li, Jiayu Chen, Maoliang Li, Xinhao Sun, Hailong Zou, Guojie Luo, and Xiang Chen. 2026. RAPID: Redundancy-Aware and Compatibility-Optimal Edge-Cloud Partitioned Infer- ence for Diverse VLA models.arXiv preprint arXiv:2603.07949(2026)

work page arXiv 2026
[45]

Gengze Zhou, Yicong Hong, and Qi Wu. 2024. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 7641–7649

2024
[46]

Junyou Zhu, Yanyuan Qiao, Siqi Zhang, Xingjian He, Qi Wu, and Jing Liu. 2025. Minivln: Efficient vision-and-language navigation by progressive knowledge distillation. In2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 97–103

2025
[47]

Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. 2023. Rt-2: Vision-language- action models transfer web knowledge to robotic control. InConference on Robot Learning. PMLR, 2165–2183

2023