arxiv: 2604.08960 · v1 · submitted 2026-04-10 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

Efficient Hierarchical Implicit Flow Q-learning for Offline Goal-conditioned Reinforcement Learning

Guoqiang Wu, Rongjian Xu, Teng Pang, Zhiqiang Dong

Pith reviewed 2026-05-10 18:08 UTC · model grok-4.3

classification 💻 cs.LG

keywords offline reinforcement learninggoal-conditioned RLhierarchical policiesmean flow policyLeJEPA lossOGBenchimplicit flow

0 comments

The pith

A goal-conditioned mean flow policy uses average velocity fields for efficient one-step sampling in hierarchical offline GCRL.

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

The paper proposes replacing Gaussian policies in hierarchical offline goal-conditioned reinforcement learning with a mean flow policy that learns an average velocity field. This field models complex target distributions for both high-level subgoal selection and low-level action generation, allowing actions to be produced in a single sampling step rather than through slower iterative methods. A companion LeJEPA loss is added to push apart goal representation embeddings during training, which produces more useful and generalizable goal encodings. The resulting method is evaluated on the OGBench suite and reports strong results on both state-vector and pixel-based tasks. A sympathetic reader would care because long-horizon goal-reaching from static offline datasets remains difficult; a practical improvement here would widen the set of real-world control problems that can be solved without online interaction or dense rewards.

Core claim

The goal-conditioned mean flow policy introduces an average velocity field into hierarchical policy modeling for offline GCRL. This captures complex target distributions for high- and low-level policies through the learned velocity field, enabling efficient action generation via one-step sampling. A LeJEPA loss repels goal representation embeddings during training to encourage more discriminative representations and improve generalization. The method achieves strong performance across both state-based and pixel-based tasks in the OGBench benchmark.

What carries the argument

The goal-conditioned mean flow policy, which learns an average velocity field to capture complex target distributions for high- and low-level policies and supports one-step sampling instead of iterative generation.

Load-bearing premise

The average velocity field learned from offline data can accurately represent the complex target distributions needed by both policy levels without producing instability or mode collapse.

What would settle it

Run the mean flow policy on a long-horizon OGBench task and check whether the one-step samples from the velocity field produce successful goal-reaching trajectories at rates comparable to or better than multi-step baselines; failure to do so on multiple seeds would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.08960 by Guoqiang Wu, Rongjian Xu, Teng Pang, Zhiqiang Dong.

**Figure 2.** Figure 2: Ablation study on the representation regularization coefficient [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗

read the original abstract

Offline goal-conditioned reinforcement learning (GCRL) is a practical reinforcement learning paradigm that aims to learn goal-conditioned policies from reward-free offline data. Despite recent advances in hierarchical architectures such as HIQL, long-horizon control in offline GCRL remains challenging due to the limited expressiveness of Gaussian policies and the inability of high-level policies to generate effective subgoals. To address these limitations, we propose the goal-conditioned mean flow policy, which introduces an average velocity field into hierarchical policy modeling for offline GCRL. Specifically, the mean flow policy captures complex target distributions for both high-level and low-level policies through a learned average velocity field, enabling efficient action generation via one-step sampling. Furthermore, considering the insufficiency of goal representation, we introduce a LeJEPA loss that repels goal representation embeddings during training, thereby encouraging more discriminative representations and improving generalization. Experimental results show that our method achieves strong performance across both state-based and pixel-based tasks in the OGBench benchmark.

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

This paper proposes a mean flow policy with average velocity field and LeJEPA loss for hierarchical offline GCRL, claiming strong OGBench results, but the abstract supplies no numbers or ablations to size the gains.

read the letter

The main point is that this work introduces a goal-conditioned mean flow policy that uses a learned average velocity field to model complex distributions for both high- and low-level policies in offline GCRL, allowing one-step sampling, and pairs it with a LeJEPA loss to encourage more discriminative goal embeddings. It does a good job calling out the shortcomings of Gaussian policies in long-horizon settings like those in HIQL and the need for better goal representations. Adapting flow concepts for average velocity to enable efficient action generation is a sensible step for offline scenarios. The reported strong performance across state and pixel tasks in OGBench suggests potential practical value for applications where new data is hard to get. The soft spots center on the missing details. The abstract mentions performance gains but gives no quantitative results, specific baselines, or ablation breakdowns, so it's not possible to see the size of the improvement or confirm the velocity field works as hoped without mode collapse. The LeJEPA loss is presented as new, but without the full paper it's hard to tell if it is distinct from existing contrastive or embedding losses in the literature. This is for people working on hierarchical and offline RL for goal-conditioned tasks, particularly those interested in efficient sampling methods. A reader in that area would find the architectural choices useful to consider. I would send it to peer review. The motivation is clear and the setting is important, so getting referee feedback on the experiments and technical details makes sense.

Referee Report

2 major / 2 minor

Summary. The paper proposes a goal-conditioned mean flow policy for hierarchical offline goal-conditioned reinforcement learning that learns an average velocity field to capture complex target distributions for high- and low-level policies, enabling efficient one-step sampling. It further introduces a LeJEPA loss to repel goal representation embeddings and improve discriminativeness. The authors claim the method achieves strong performance on both state-based and pixel-based tasks in the OGBench benchmark.

Significance. If the empirical results hold, the work could meaningfully advance offline GCRL by addressing expressiveness limits of Gaussian policies and weak goal representations in hierarchical settings. The mean-flow formulation for one-step sampling from complex distributions is a potentially useful idea for long-horizon control from reward-free data.

major comments (2)

Abstract: the central claim that the method 'achieves strong performance across both state-based and pixel-based tasks' is stated without any quantitative results, baselines, metrics, or ablation details. This is load-bearing for an empirical contribution and prevents verification of the performance gains.
The provided text supplies no equations or algorithmic pseudocode for the average velocity field or the LeJEPA loss, so it is impossible to check whether the velocity field is learned in a way that actually avoids mode collapse or instability on offline data (the weakest assumption identified in the review).

minor comments (2)

Abstract: the acronym LeJEPA is introduced without expansion or a one-sentence description of what the loss does.
The manuscript would benefit from a short related-work paragraph contrasting the mean-flow policy with prior implicit or flow-based policies in GCRL.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We agree that the abstract requires more concrete details to support our claims, and we will ensure the technical components are presented with full clarity, including equations and pseudocode.

read point-by-point responses

Referee: Abstract: the central claim that the method 'achieves strong performance across both state-based and pixel-based tasks' is stated without any quantitative results, baselines, metrics, or ablation details. This is load-bearing for an empirical contribution and prevents verification of the performance gains.

Authors: We agree that the abstract should be more informative. In the revision, we will incorporate specific quantitative results drawn from our OGBench experiments, including average success rates on state-based and pixel-based tasks, direct comparisons to baselines such as HIQL, and references to key metrics and ablations from the results section. This will make the performance claims verifiable at a glance. revision: yes
Referee: The provided text supplies no equations or algorithmic pseudocode for the average velocity field or the LeJEPA loss, so it is impossible to check whether the velocity field is learned in a way that actually avoids mode collapse or instability on offline data (the weakest assumption identified in the review).

Authors: The full manuscript contains the mathematical definitions of the mean flow policy (including the average velocity field) in Section 3.2 and the LeJEPA loss in Section 3.3. However, to improve accessibility, we will add explicit algorithmic pseudocode for training and one-step sampling in the revised version (main text or appendix). We will also expand the discussion on how the formulation helps capture complex distributions and reduces mode collapse risks in the offline hierarchical setting. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes a new hierarchical architecture consisting of a goal-conditioned mean flow policy (with learned average velocity field for one-step sampling) and a LeJEPA loss for goal embeddings. These are presented as architectural innovations for offline GCRL, supported by empirical results on OGBench. No load-bearing equations, predictions, or uniqueness claims reduce by construction to fitted parameters, self-definitions, or self-citation chains. The derivation remains self-contained and independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

Abstract-only review provides no explicit free parameters, axioms, or independent evidence for new entities; the velocity field and LeJEPA loss are introduced without stated assumptions or external validation.

invented entities (2)

goal-conditioned mean flow policy no independent evidence
purpose: Captures complex target distributions for high- and low-level policies via learned average velocity field for one-step sampling
Introduced to address limited expressiveness of Gaussian policies and ineffective subgoal generation
LeJEPA loss no independent evidence
purpose: Repels goal representation embeddings to encourage discriminative representations and better generalization
Proposed to fix insufficiency of goal representation in hierarchical policies

pith-pipeline@v0.9.0 · 5470 in / 1143 out tokens · 37837 ms · 2026-05-10T18:08:02.164440+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

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

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
the mean flow policy captures complex target distributions ... through a learned average velocity field, enabling efficient action generation via one-step sampling
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean absolute_floor_iff_bare_distinguishability unclear
LeJEPA loss that repels goal representation embeddings ... SIGReg ... isotropic Gaussian

Reference graph

Works this paper leans on

45 extracted references · 24 canonical work pages · 11 internal anchors

[1]

Learning to achieve goals

Leslie Pack Kaelbling. Learning to achieve goals. InIJCAI, volume 2, pages 1094–8, 1993

1993
[2]

Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems

Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforce- ment learning: Tutorial, review, and perspectives on open problems.arXiv preprint arXiv:2005.01643, 2020

work page internal anchor Pith review arXiv 2005
[3]

Ogbench: Benchmarking offline goal-conditioned rl.arXiv preprint arXiv:2410.20092,

Seohong Park, Kevin Frans, Benjamin Eysenbach, and Sergey Levine. Ogbench: Benchmarking offline goal-conditioned rl.arXiv preprint arXiv:2410.20092, 2024

work page arXiv 2024
[4]

Hierarchi- cal reinforcement learning: A comprehensive survey.ACM Computing Surveys (CSUR), 54(5):1–35, 2021

Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchi- cal reinforcement learning: A comprehensive survey.ACM Computing Surveys (CSUR), 54(5):1–35, 2021

2021
[5]

Landmark-guided subgoal gen- eration in hierarchical reinforcement learning.Advances in neural information processing systems, 34:28336–28349, 2021

Junsu Kim, Younggyo Seo, and Jinwoo Shin. Landmark-guided subgoal gen- eration in hierarchical reinforcement learning.Advances in neural information processing systems, 34:28336–28349, 2021. 14

2021
[6]

Hiql: Offline goal-conditioned rl with latent states as actions.Advances in Neural In- formation Processing Systems, 36:34866–34891, 2023

Seohong Park, Dibya Ghosh, Benjamin Eysenbach, and Sergey Levine. Hiql: Offline goal-conditioned rl with latent states as actions.Advances in Neural In- formation Processing Systems, 36:34866–34891, 2023

2023
[7]

What Matters in Learning from Offline Human Demonstrations for Robot Manipulation

Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriany, Chen Wang, Ro- hun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Mart´ın-Mart´ın. What matters in learning from offline human demonstrations for robot manipula- tion.arXiv preprint arXiv:2108.03298, 2021

work page internal anchor Pith review arXiv 2021
[8]

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

Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6892–6903. IEEE, 2024

2024
[9]

Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

2020
[10]

Flow Matching Guide and Code

Yaron Lipman, Marton Havasi, Peter Holderrieth, Neta Shaul, Matt Le, Brian Karrer, Ricky TQ Chen, David Lopez-Paz, Heli Ben-Hamu, and Itai Gat. Flow matching guide and code.arXiv preprint arXiv:2412.06264, 2024

work page internal anchor Pith review arXiv 2024
[11]

Is conditional generative modeling all you need for decision-making? arXiv preprint arXiv:2211.15657, 2022

Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision- making?arXiv preprint arXiv:2211.15657, 2022

work page arXiv 2022
[12]

AdaptDiffuser: Diffusion models as adaptive self-evolving planners.arXiv preprint arXiv:2302.01877, 2023

Zhixuan Liang, Yao Mu, Mingyu Ding, Fei Ni, Masayoshi Tomizuka, and Ping Luo. Adaptdiffuser: Diffusion models as adaptive self-evolving planners.arXiv preprint arXiv:2302.01877, 2023

work page arXiv 2023
[13]

Diffusion policies as an expressive policy class for offline reinforcement learning.arXiv preprint arXiv:2208.06193,

Zhendong Wang, Jonathan J Hunt, and Mingyuan Zhou. Diffusion policies as an expressive policy class for offline reinforcement learning.arXiv preprint arXiv:2208.06193, 2022

work page arXiv 2022
[14]

Diffusion policy: Visuomotor pol- icy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor pol- icy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

2025
[15]

Flow Q - Learning , May 2025 c

Seohong Park, Qiyang Li, and Sergey Levine. Flow q-learning.arXiv preprint arXiv:2502.02538, 2025

work page arXiv 2025
[16]

Efficient diffusion policies for offline reinforcement learning.Advances in Neural Infor- mation Processing Systems, 36:67195–67212, 2023

Bingyi Kang, Xiao Ma, Chao Du, Tianyu Pang, and Shuicheng Yan. Efficient diffusion policies for offline reinforcement learning.Advances in Neural Infor- mation Processing Systems, 36:67195–67212, 2023

2023
[17]

Mean Flows for One-step Generative Modeling

Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling.arXiv preprint arXiv:2505.13447, 2025. 15

work page internal anchor Pith review arXiv 2025
[18]

Learning to reach goals via iterated supervised learning, 2020

Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Ben- jamin Eysenbach, and Sergey Levine. Learning to reach goals via iterated super- vised learning.arXiv preprint arXiv:1912.06088, 2019

work page arXiv 1912
[19]

Contrastive learning as goal-conditioned reinforcement learning.Advances in Neural Information Processing Systems, 35:35603–35620, 2022

Benjamin Eysenbach, Tianjun Zhang, Sergey Levine, and Russ R Salakhutdinov. Contrastive learning as goal-conditioned reinforcement learning.Advances in Neural Information Processing Systems, 35:35603–35620, 2022

2022
[20]

Mapping state space using landmarks for universal goal reaching.Advances in Neural Information Processing Systems, 32, 2019

Zhiao Huang, Fangchen Liu, and Hao Su. Mapping state space using landmarks for universal goal reaching.Advances in Neural Information Processing Systems, 32, 2019

2019
[21]

Offline Reinforcement Learning with Implicit Q-Learning

Ilya Kostrikov, Ashvin Nair, and Sergey Levine. Offline reinforcement learning with implicit q-learning.arXiv preprint arXiv:2110.06169, 2021

work page internal anchor Pith review arXiv 2021
[22]

Advantage-Weighted Regression: Simple and Scalable Off-Policy Reinforcement Learning

Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage- weighted regression: Simple and scalable off-policy reinforcement learning. arXiv preprint arXiv:1910.00177, 2019

work page internal anchor Pith review arXiv 1910
[23]

Flow Matching for Generative Modeling

Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[24]

Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[25]

Building Normalizing Flows with Stochastic Interpolants

Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571, 2022

work page internal anchor Pith review arXiv 2022
[26]

Option-aware temporally abstracted value for offline goal-conditioned reinforcement learning

Hongjoon Ahn, Heewoong Choi, Jisu Han, and Taesup Moon. Option-aware temporally abstracted value for offline goal-conditioned reinforcement learning. arXiv preprint arXiv:2505.12737, 2025

work page arXiv 2025
[27]

LeJEPA: Provable and scalable self-supervised learning without the heuristics.arXiv preprint arXiv:2511.08544, 2025

Randall Balestriero and Yann LeCun. Lejepa: Provable and scalable self- supervised learning without the heuristics.arXiv preprint arXiv:2511.08544, 2025

work page arXiv 2025
[28]

Signature verification using a” siamese” time delay neural network.Ad- vances in neural information processing systems, 6, 1993

Jane Bromley, Isabelle Guyon, Yann LeCun, Eduard S ¨ackinger, and Roopak Shah. Signature verification using a” siamese” time delay neural network.Ad- vances in neural information processing systems, 6, 1993

1993
[29]

A path towards autonomous machine intelligence version 0.9

Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27.Open Review, 62(1):1–62, 2022

2022
[30]

Tutorial on joint embedding predictive architectures (jepa): Founda- tions, applications, and future directions.Authorea Preprints, 2025

Mehdi Monemi, Maryam Chinipardaz, Mehdi Rasti, Mehdi Bennis, and Matti Latva-Aho. Tutorial on joint embedding predictive architectures (jepa): Founda- tions, applications, and future directions.Authorea Preprints, 2025. 16

2025
[31]

A sim- ple framework for contrastive learning of visual representations

Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A sim- ple framework for contrastive learning of visual representations. InInternational conference on machine learning, pages 1597–1607. PmLR, 2020

2020
[32]

An empirical study of training self- supervised vision transformers

Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self- supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 9640–9649, 2021

2021
[33]

Goal-conditioned re- inforcement learning: Problems and solutions.arXiv preprint arXiv:2201.08299,

Minghuan Liu, Menghui Zhu, and Weinan Zhang. Goal-conditioned reinforce- ment learning: Problems and solutions.arXiv preprint arXiv:2201.08299, 2022

work page arXiv 2022
[34]

Optimal goal- reaching reinforcement learning via quasimetric learning

Tongzhou Wang, Antonio Torralba, Phillip Isola, and Amy Zhang. Optimal goal- reaching reinforcement learning via quasimetric learning. InInternational Con- ference on Machine Learning, pages 36411–36430. PMLR, 2023

2023
[35]

Learning temporal distances: Contrastive successor features can provide a metric structure for decision-making.arXiv preprint arXiv:2406.17098, 2024

Vivek Myers, Chongyi Zheng, Anca Dragan, Sergey Levine, and Benjamin Ey- senbach. Learning temporal distances: Contrastive successor features can provide a metric structure for decision-making.arXiv preprint arXiv:2406.17098, 2024

work page arXiv 2024
[36]

arXiv preprint arXiv:2210.00030 , year=

Yecheng Jason Ma, Shagun Sodhani, Dinesh Jayaraman, Osbert Bastani, Vikash Kumar, and Amy Zhang. Vip: Towards universal visual reward and representation via value-implicit pre-training.arXiv preprint arXiv:2210.00030, 2022

work page arXiv 2022
[37]

Grace Liu, Michael Tang, and Benjamin Eysenbach

Grace Liu, Michael Tang, and Benjamin Eysenbach. A single goal is all you need: Skills and exploration emerge from contrastive rl without rewards, demon- strations, or subgoals.arXiv preprint arXiv:2408.05804, 2024

work page arXiv 2024
[38]

Deep unsupervised learning using nonequilibrium thermodynamics

Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInterna- tional conference on machine learning, pages 2256–2265. pmlr, 2015

2015
[39]

Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

2021
[40]

Scal- ing rectified flow transformers for high-resolution image synthesis

Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scal- ing rectified flow transformers for high-resolution image synthesis. InForty-first international conference on machine learning, 2024

2024
[41]

Planning with Diffusion for Flexible Behavior Synthesis

Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis.arXiv preprint arXiv:2205.09991, 2022

work page internal anchor Pith review arXiv 2022
[42]

Consistency mod- els

Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency mod- els. 2023

2023
[43]

Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models

Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024. 17

work page internal anchor Pith review arXiv 2024
[44]

Consistency models made easy

Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy.arXiv preprint arXiv:2406.14548, 2024

work page arXiv 2024
[45]

A minimalist approach to offline reinforcement learning.Advances in neural information processing systems, 34:20132–20145, 2021

Scott Fujimoto and Shixiang Shane Gu. A minimalist approach to offline reinforcement learning.Advances in neural information processing systems, 34:20132–20145, 2021. 18

2021