Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Extremum Flow Matching for Offline Goal Conditioned Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Eestimating the minimum of a distribution by reading the edge of a flow-matching source distribution lets goal-conditioned policies pick optimal paths from uncurated play data.

desk verdict Clever boundary-reading trick for return-conditioned offline RL, backed by broad experiments, but the central theoretical claim ignores the discreteness of returns and needs tightening. read the letter →

arxiv 2505.19717 v2 pith:O43DG7NN submitted 2025-05-26 cs.RO

Quentin Rouxel (CUHK) , Clemente Donoso , Fei Chen (CUHK) , Serena Ivaldi , Jean-Baptiste Mouret This is my paper · ORCID
classification cs.RO
keywords extremumflowmatchinggoal-conditionedreinforcementlearningofflineimitationfromplayreturnconditioninghumanoidmanipulationtrajectorystitching
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

The paper introduces Extremum Flow Matching, a way to read the minimum (or maximum) of a learned distribution from the boundary of a uniform source distribution in Flow Matching. In one dimension, the flow's deterministic, non-intersecting paths map the source endpoints to the target support's endpoints, so evaluating the flow at eps=0 gives the smallest value in the target. The paper uses this to estimate the shortest demonstrated time-distance to a goal from play data, then conditions a goal-conditioned policy on that estimated optimal return. The authors build four flow-matching agent architectures (critic, actor, planner, world model), add a Bellman-style backup for stitching trajectory segments, and evaluate on OGBench, a planar pushing task with different demonstration behaviors, and a real Talos humanoid performing kitchen manipulation from image goals.

What carries the argument

The load-bearing object is the boundary-mapping property of a flow trained with a uniform source: because flow paths are continuous and non-intersecting, the endpoints 0 and 1 of the source interval are transported to the minimum and maximum of the target distribution's support. The Critic is a flow-matching model F : U(0,1)|(o,g) -> d trained with the squared regression loss on interpolated returns, and evaluated at eps = 0 at inference to yield the minimal demonstrated time-distance to the goal.

What would settle it

On a synthetic conditional distribution whose true minimum as a function of context c is known analytically, train the flow-matching critic with fixed capacity and data, then compare F(0|c) against the true minimum across many contexts; if the error is large near sharp or sparse-tail regions, the boundary estimator fails. Alternatively, in a gridworld where the shortest path length between any two states is computable, check whether Critic(0|o,g) equals the true shortest distance; systematic overestimation would indicate the flow does not reach the support boundary.

Watch

Extended reading notes

Core claim

The paper claims that Flow Matching's deterministic transport and ability to accept any source distribution turn the problem of estimating a distribution's extrema into a single evaluation of the learned flow at the source boundary. For a target distribution conditioned on context c, training F : U(0,1)|c -> x(c) with the standard flow-matching regression makes F(0|c) and F(1|c) the minimal and maximal values of the conditional support. In a multi-dimensional setting, the paper decomposes the generative process into a one-dimensional model for the axis of interest and a conditional model for the remaining coordinates, so the extremal value along one axis can be sampled with the corresponding typical other coordinates. The authors then define Critic(0|o,g) as the estimated shortest distance between observation o and goal g, and use this in-distribution value to condition actors and planners, obtaining goal-reaching policies from unstructured play.

Load-bearing premise

The boundary-mapping property is guaranteed for an exactly learned homeomorphic flow, but the neural flow is trained with a squared-error regression on interior interpolated samples, so the source endpoints 0 and 1 are never supervised directly; if the learned vector field deviates there, Critic(0|o,g) will be a biased estimate of the true minimum return.

Editorial extensions

If this is right

  • Policy conditioning on the estimated shortest distance extracts near-optimal behavior from datasets where no trajectory was optimal.
  • Because the conditioning value comes from the learned distribution's boundary, it stays in-distribution, avoiding the out-of-distribution extrapolation of hand-tuned extreme returns.
  • The recursive bootstrap in Eq. 5 stitches together returns from disjoint episodes, enabling long-horizon tasks that were never demonstrated in one trajectory.
  • The same flow-matching critic can score candidate plans in rejection-sampling agents without any reward engineering.
  • Play data alone, without expert demonstrations, can produce multi-step vision-based manipulation on a real humanoid.

Reading between the lines

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

  • The extremum reading is not limited to the min/max; evaluating the flow at intermediate eps values yields quantile-like estimates of the learned distribution, which could be used to interpolate between conservative and optimistic conditioning.
  • The method transfers to other deterministic transport models as long as the source has known boundaries, since only the non-intersection of paths is required.
  • The boundary estimates are only as good as the flow at the endpoints; explicitly supervising the vector field at eps=0 and eps=1, or adding boundary-aware sampling, could remove the main bias source.
  • The observed trade-off where RL backups hurt expert data but help stitched play suggests the backup's noise interacts with the boundary estimate; a testable prediction is that the degradation grows with the number of backup steps.
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

4 major / 5 minor

Summary. The paper proposes Extremum Flow Matching, a method to estimate the minimum or maximum of a conditional distribution by exploiting the deterministic transport property of flow matching: with a uniform source distribution, the boundary points of the source are claimed to map to the boundary of the target distribution. The method is applied to offline goal-conditioned imitation and reinforcement learning from play data, yielding several agents (FM-GC, FM-AC, FM-PC, FM-PS, FM-AS) with optional recursive bootstrapping. The agents are evaluated on the OGBench benchmark, on a planar pushing task with different demonstration behaviors, and on a real humanoid robot (Talos) performing kitchen manipulation from image observations.

Significance. The core idea of using flow-matching boundaries as distributional extrema estimates is novel and offers a practical alternative to expectile regression, with the advantage of selecting the conditioning quantile at inference. The paper is experimentally rich: it compares multiple agent architectures, analyzes the effect of data-collection behavior, and includes a real-robot validation on a full-size humanoid, which is commendable and rare. The authors provide code and videos, supporting reproducibility. However, the theoretical foundation is not fully established for the discrete return distributions that are central to the problem setting, which casts doubt on the central optimality mechanism. If the discrete-distribution issue is addressed, the work could be a meaningful contribution to goal-conditioned offline RL.

major comments (4)
  1. [III-B, II-B] The boundary-mapping property stated in Section III-B (and Fig. 1) assumes an exact homeomorphic flow from a continuous uniform source to the target distribution. However, the paper's own formalism (Section II-B) defines returns d as integer time-step distances, so the target distribution of returns is discrete. A continuous deterministic flow cannot transport a continuous uniform distribution exactly to a discrete empirical distribution; the learned Critic at epsilon=0 is the left tail of a smoothed approximation (e.g., a mixture of Gaussians), not the true minimum of the demonstrated returns. Therefore Critic(0|o,g) may be systematically below the smallest observed d, directly contradicting the in-distribution conditioning claim in Section III-C. This is a structural bias, not merely a regression error, and it affects all agents that use Critic(0|o,g) as the optimal return (FM-AC, FM-PC, and the RL variants). The paper does not test this scenario; the illustration in Fig. 3 uses a continuous multimodal distribution, not a discrete return distribution. Please either provide a rigorous analysis of the bias for discrete targets, or modify the extremum estimator (e.g., using a quantile or a post-hoc correction) and re-evaluate.
  2. [III-B, Eqs. (3)-(4)] The multi-dimensional extension relies on the assumption that F1 (the flow for the return dimension z) produces the exact extremum of P(z). While the decomposition P(x)=P(z)P(y|z) is mathematically sound, the paper provides no proof that the flow trained with the squared-error loss in Eq. (1) maps the source boundary to the target extremum, even in the continuous case. The claim that 'flow paths do not intersect' implies order preservation only for an exact monotone homeomorphism; a neural-network vector field trained by regression is not guaranteed to be monotone or to map the boundary exactly. Please provide a formal statement of the conditions under which the extremum estimate is unbiased, or an empirical validation on distributions with known extrema (including discrete targets). Without this, the optimality mechanism of the proposed agents is not justified.
  3. [IV-B, Eq. (5)] The recursive backup in Eq. (5) introduces two free parameters (rg and the distribution of epsilon_g) and is described as a 'regularization hyperparameter to mitigate underestimation bias.' However, the experiments show a dramatic performance drop on the expert dataset when the RL backup is used (e.g., FM-AC-use-RL 32±8 vs. FM-AC-no-RL 96±5 in Fig. 7), and the paper explicitly states that 'the underlying cause of this trade-off remains unclear.' This is a significant unexplained failure mode: the proposed regularization does not behave as intended on at least one dataset, which undermines the claim that the backup is a principled component. The authors should analyze this failure mode (e.g., whether the double-network trick is insufficient, or whether the random epsilon_g draws push returns out of distribution) and provide a more robust backup mechanism or a clear caveat about when use-RL should be applied.
  4. [IV-A] The comparison with OGBench baselines is asymmetric: all proposed agents are trained with a single fixed set of hyperparameters across all tasks, whereas the OGBench results use per-task tuned hyperparameters for each baseline. This makes the reported performance differences difficult to interpret. For a fair comparison, the authors should either tune their own hyperparameters per task (and report the tuning procedure) or use the same fixed-hyperparameter protocol for the baselines. At minimum, the paper should discuss how the fixed hyperparameters were chosen and whether the conclusions change under a different hyperparameter setting.
minor comments (5)
  1. [II-B] The notation 'd ∼ U(0, Lg)' suggests a continuous uniform distribution, but the text states that d is in N (natural numbers). Please use a discrete uniform notation, e.g., d ∼ Uniform({0,1,...,Lg}).
  2. [II-A, Eq. (1)] The flow matching loss in Eq. (1) is written without an explicit expectation over data and time. Adding the expectation (e.g., E_{x_dst,x_src,t}) would improve clarity.
  3. [Fig. 5] The results table in Fig. 5 is dense and difficult to parse. Consider highlighting the best result in each row (e.g., bold) and adding error bars or significance tests where possible.
  4. [IV-B] The paper mentions that FM-AS's performance 'consistently degraded over the course of training' but does not provide a learning curve or analysis. Including a plot and a brief discussion of this instability would be helpful, as it suggests a training issue that could affect the interpretation of the world-model agent.
  5. [V] The limitations section notes that the proposed agents require significantly more training time, but no quantitative comparison is given. Adding wall-clock training times relative to the OGBench baselines would make the trade-off concrete.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the extremum readout is an inference-time evaluation of a flow model trained on the return distribution, and the RL bootstrap is a standard TD-style self-targeting mechanism, not a reduction of the central claim.

full rationale

The claimed derivation chain is not circular. In Section III-B, the extremum estimate Critic(0|o,g) = F(0|o,g) is obtained by training the flow model F on the empirical return distribution P(d|o,g) with the conditional Flow Matching loss (Eq. 1); the uniform source is not fitted to the minimum, and the endpoint evaluation is an inference-time operation of the learned transport, so the minimum is read from the data rather than imposed by definition. The boundary-mapping property, stated as 'flow paths do not intersect', is an external mathematical property of deterministic ODE flows, and the paper does not import it through a self-citation or an author-specific uniqueness theorem. Eq. (5) does use the Critic to produce its own augmented return targets, but this is a standard Bellman-style bootstrapping device confined to the use-RL variants; the no-RL agents are evaluated independently, and the central extremum method does not reduce to this backup. Self-citations [29], [44]-[46] support generic claims about flow matching for imitation and the robot controller, and are not load-bearing for the extremum claim. The experiments against external OGBench baselines and the real-robot deployment give independent empirical content. The discrete-return concern raised by the skeptic is a correctness and robustness gap, not a circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the boundary-preserving property of an idealized flow, on the accuracy of the marginal/conditional decomposition, and on the identification of optimality with the shortest demonstrated return. These are modeling assumptions, not fitted numbers; the only explicit free hyperparameter is the bootstrap regularization rg.

free parameters (2)
  • rg
    Regularization scaling factor in the recursive bootstrap (Eq. 5), chosen from [0,1] to control under-estimation bias; value not specified in the paper.
  • Lg
    Maximum goal horizon for hindsight goal sampling in Section II-B; controls the range of returns and is a free design choice.
assumptions (3)
  • domain assumption The learned flow matching model is a homeomorphism between source and target supports, preserving boundary correspondence.
    Section III-B relies on this to map the uniform source endpoints to distributional extrema; the regression loss gives no boundary guarantee.
  • domain assumption The decomposition P(z,y)=P(z)P(y|z) and separate flow models let the extremal z sample from the correct conditional y.
    Equations (3)-(4) and Fig. 4 assume the marginal and conditional flows are accurate enough for the extremal conditioning to be valid.
  • domain assumption The shortest demonstrated distance to a goal is the appropriate optimality target for offline goal-conditioned RL.
    Section III-A defines optimality as min P(d|o,g); this is the best the dataset can certify, not necessarily the true optimal return.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extremum Flow Matching for Offline Goal Conditioned Reinforcement Learning." pith.science (2026). https://pith.science/paper/O43DG7NN

@misc{pith2026250519717,
  author       = {Pith},
  title        = {Pith review of: Extremum Flow Matching for Offline Goal Conditioned Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O43DG7NN}},
  note         = {Machine review of arXiv:2505.19717}
}
read the original abstract

Imitation learning is a promising approach for enabling generalist capabilities in humanoid robots, but its scaling is fundamentally constrained by the scarcity of high-quality expert demonstrations. This limitation can be mitigated by leveraging suboptimal, open-ended play data, often easier to collect and offering greater diversity. This work builds upon recent advances in generative modeling, specifically Flow Matching, an alternative to Diffusion models. We introduce a method for estimating the minimum or maximum of the learned distribution by leveraging the unique properties of Flow Matching, namely, deterministic transport and support for arbitrary source distributions. We apply this method to develop several goal-conditioned imitation and reinforcement learning algorithms based on Flow Matching, where policies are conditioned on both current and goal observations. We explore and compare different architectural configurations by combining core components, such as critic, planner, actor, or world model, in various ways. We evaluated our agents on the OGBench benchmark and analyzed how different demonstration behaviors during data collection affect performance in a 2D non-prehensile pushing task. Furthermore, we validated our approach on real hardware by deploying it on the Talos humanoid robot to perform complex manipulation tasks based on high-dimensional image observations, featuring a sequence of pick-and-place and articulated object manipulation in a realistic kitchen environment. Experimental videos and code are available at: https://hucebot.github.io/extremum_flow_matching_website/

Figures

Figures reproduced from arXiv: 2505.19717 by the authors.

Figure 1
Figure 1. Goal conditioned policy using Extremum Flow Matching for manip [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Imitation and offline reinforcement learning from play data face [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Comparison between Expectile Regression and Extremum Flow [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Comparison of our proposed agents on the OGBench benchmark [40]. We report the average of binary success rates (%) across state-based [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Planar pushing task in a maze (top left): the agent controls the small [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Success rates of proposed agents across three datasets with varying demonstration behaviors. Performance is evaluated on the same set of [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Kitchen task setup for image-based experiments on the Talos [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: When conditioned on the current and desired goal images shown in Fig. 1 (top), the [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Native Extrapolation Awareness in Flow-Based Conditional Generation

    cs.LG 2026-02 conditional novelty 6.0 of 10

    A contrastive flow-matching objective makes off-manifold conditions produce curved trajectories, so path curvature (the DOT score) separates invalid from valid inputs.

Reference graph

Works this paper leans on

48 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    Implicit behavioral cloning,

    P. Florence et al. , “Implicit behavioral cloning,” in CoRL, 2022

    work page 2022

  2. [2]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi et al. , “Diffusion policy: Visuomotor policy learning via action diffusion,” IJRR, 2023

    work page 2023

  3. [3]

    π0: A vision-language-action flow model for general robot control,

    K. Black et al. , “ π0: A vision-language-action flow model for general robot control,” arXiv preprint arXiv:2410.24164 , 2024

    arXiv 2024

  4. [4]

    Goal-conditioned imitation learning,

    Y . Ding, C. Florensa, P. Abbeel, and M. Phielipp, “Goal-conditioned imitation learning,” NeurIPS, 2019

    work page 2019

  5. [5]

    Goal-conditioned imitation learning using score-based diffusion policies,

    M. Reuss, M. Li, X. Jia, and R. Lioutikov, “Goal-conditioned imitation learning using score-based diffusion policies,” RSS, 2023

    work page 2023

  6. [6]

    From play to policy: Conditional behavior generation from uncurated robot data,

    Z. J. Cui et al. , “From play to policy: Conditional behavior generation from uncurated robot data,” ICLR, 2023

    work page 2023

  7. [7]

    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,” in IEEE ICRA, 2024

    work page 2024

  8. [8]

    Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,

    M. Reuss et al. , “Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,” RSS, 2024

    work page 2024

Show all 48 references
  1. [9]

    Ai robots and humanoid ai: Review, perspectives and directions,

    L. Cao, “Ai robots and humanoid ai: Review, perspectives and directions,” arXiv preprint arXiv:2405.15775 , 2024

    2024

  2. [10]

    Advancements in humanoid robots: A comprehensive review and future prospects,

    Y . Tong, H. Liu, and Z. Zhang, “Advancements in humanoid robots: A comprehensive review and future prospects,” IEEE/CAA Journal of Automatica Sinica , vol. 11, no. 2, pp. 301–328, 2024

    2024

  3. [11]

    Learning latent plans from play,

    C. Lynch, M. Khansari, T. Xiao, V . Kumar, J. Tompson, S. Levine, and P. Sermanet, “Learning latent plans from play,” in CoRL, 2020

    2020

  4. [12]

    Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks,

    O. Mees, L. Hermann, E. Rosete-Beas, and W. Burgard, “Calvin: A benchmark for language-conditioned policy learning for long-horizon robot manipulation tasks,” IEEE RA-L , 2022

    2022

  5. [13]

    Mimicplay: Long-horizon imitation learning by watching human play,

    C. Wang et al. , “Mimicplay: Long-horizon imitation learning by watching human play,” CoRL, 2023

    2023

  6. [14]

    Is conditional generative modeling all you need for decision-making?

    A. Ajay et al. , “Is conditional generative modeling all you need for decision-making?” ICLR, 2023

    2023

  7. [15]

    Offline reinforcement learning with implicit q-learning,

    I. Kostrikov, A. Nair, and S. Levine, “Offline reinforcement learning with implicit q-learning,” ICLR, 2021. Accepted Preprint at 2025 IEEE-RAS 24th International Conference on Humanoid Robots (Humanoids), August 2025 Fig. 9. When conditioned on the current and desired goal im...

    2021

  8. [16]

    Enhancing decision transformer with diffusion-based trajectory branch generation,

    Z. Liu et al. , “Enhancing decision transformer with diffusion-based trajectory branch generation,” arXiv preprint arXiv:2411.11327 , 2024

    2024 arXiv

  9. [17]

    Hiql: Offline goal-conditioned rl with latent states as actions,

    S. Park, D. Ghosh, B. Eysenbach, and S. Levine, “Hiql: Offline goal-conditioned rl with latent states as actions,” NeurIPS, 2023

    2023

  10. [18]

    Stitching sub-trajectories with conditional diffusion model for goal-conditioned offline rl,

    S. Kim et al. , “Stitching sub-trajectories with conditional diffusion model for goal-conditioned offline rl,” in AAAI, 2024

    2024

  11. [19]

    What makes a good diffusion planner for decision making?

    H. Lu, D. Han, Y . Shen, and D. Li, “What makes a good diffusion planner for decision making?” in ICLR, 2025

    2025

  12. [20]

    Denoising diffusion probabilistic models,

    J. Ho et al. , “Denoising diffusion probabilistic models,” in NeurIPS, 2020

    2020

  13. [21]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in ICLR, 2021

    2021

  14. [22]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and qiang liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” in ICLR, 2023

    2023

  15. [23]

    Planning with diffusion for flexible behavior synthesis,

    M. Janner, Y . Du, J. B. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” ICML, 2022

    2022

  16. [24]

    Generative skill chaining: Long-horizon skill planning with diffusion models,

    U. A. Mishra, S. Xue, Y . Chen, and D. Xu, “Generative skill chaining: Long-horizon skill planning with diffusion models,” in CoRL, 2023

    2023

  17. [25]

    Simple hierarchical planning with diffusion,

    C. Chen, F. Deng, K. Kawaguchi, C. Gulcehre, and S. Ahn, “Simple hierarchical planning with diffusion,” ICLR, 2024

    2024

  18. [26]

    Generative trajectory stitching through diffusion composition,

    Y . Luo, U. A. Mishra, Y . Du, and D. Xu, “Generative trajectory stitching through diffusion composition,” arXiv preprint arXiv:2503.05153, 2025

    2025 arXiv

  19. [27]

    Stochastic interpolants: A unifying framework for flows and diffusions,

    M. S. Albergo et al. , “Stochastic interpolants: A unifying framework for flows and diffusions,” arXiv preprint arXiv:2303.08797 , 2023

    2023 arXiv

  20. [28]

    Building normalizing flows with stochastic interpolants,

    M. S. Albergo and E. Vanden-Eijnden, “Building normalizing flows with stochastic interpolants,” in ICLR, 2023

    2023

  21. [29]

    Flow matching imitation learning for multi-support manipulation,

    Q. Rouxel et al. , “Flow matching imitation learning for multi-support manipulation,” in IEEE-RAS Humanoids , 2024

    2024

  22. [30]

    Adaflow: Imitation learning with variance-adaptive flow-based policies,

    X. Hu, Q. Liu, X. Liu, and B. Liu, “Adaflow: Imitation learning with variance-adaptive flow-based policies,” NeurIPS, 2024

    2024

  23. [31]

    Riemannian flow matching policy for robot motion learning,

    M. Braun, N. Jaquier, L. Rozo, and T. Asfour, “Riemannian flow matching policy for robot motion learning,” in IEEE/RSJ IROS, 2024

    2024

  24. [32]

    Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation,

    Q. Zhang, Z. Liu, H. Fan, G. Liu, B. Zeng, and S. Liu, “Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation,” in AAAI, 2025

    2025

  25. [33]

    Energy-weighted flow matching for offline reinforcement learning,

    S. Zhang, W. Zhang, and Q. Gu, “Energy-weighted flow matching for offline reinforcement learning,” arXiv preprint arXiv:2503.04975 , 2025

    2025 arXiv

  26. [34]

    Flow q-learning,

    S. Park, Q. Li, and S. Levine, “Flow q-learning,” arXiv preprint arXiv:2502.02538, 2025

    2025 arXiv

  27. [35]

    Hindsight experience replay,

    M. Andrychowicz et al., “Hindsight experience replay,” NeurIPS, 2017

    2017

  28. [36]

    Asymmetric least squares estimation and testing,

    W. K. Newey and J. L. Powell, “Asymmetric least squares estimation and testing,” Econometrica: Journal of the Econometric Society , 1987

    1987

  29. [37]

    Diffusion world model: Future modeling beyond step-by-step rollout for offline reinforcement learning,

    Z. Ding, A. Zhang, Y . Tian, and Q. Zheng, “Diffusion world model: Future modeling beyond step-by-step rollout for offline reinforcement learning,” arXiv preprint arXiv:2402.03570 , 2024

    2024 arXiv

  30. [38]

    Learning from reward-free offline data: A case for planning with latent dynamics models,

    V . Sobal, W. Zhang, K. Cho, R. Balestriero, T. G. Rudner, and Y . LeCun, “Learning from reward-free offline data: A case for planning with latent dynamics models,” arXiv preprint arXiv:2502.14819 , 2025

    2025

  31. [39]

    Double q-learning,

    H. Hasselt, “Double q-learning,” NeurIPS, 2010

    2010

  32. [40]

    Ogbench: Benchmarking offline goal-conditioned rl,

    S. Park, K. Frans, B. Eysenbach, and S. Levine, “Ogbench: Benchmarking offline goal-conditioned rl,” in ICLR, 2025

    2025

  33. [41]

    Learning to reach goals via iterated supervised learning,

    D. Ghosh et al. , “Learning to reach goals via iterated supervised learning,” ICLR, 2021

    2021

  34. [42]

    Optimal goal-reaching reinforcement learning via quasimetric learning,

    T. Wang, A. Torralba, P. Isola, and A. Zhang, “Optimal goal-reaching reinforcement learning via quasimetric learning,” in ICML, 2023

    2023

  35. [43]

    Contrastive learning as goal-conditioned reinforcement learning,

    B. Eysenbach et al. , “Contrastive learning as goal-conditioned reinforcement learning,” NeurIPS, 2022

    2022

  36. [44]

    Multi-contact whole-body force control for position-controlled robots,

    Q. Rouxel, S. Ivaldi, and J.-B. Mouret, “Multi-contact whole-body force control for position-controlled robots,” IEEE RA-L , 2024

    2024

  37. [45]

    Multicontact motion retarget- ing using whole-body optimization of full kinematics and sequential force equilibrium,

    Q. Rouxel, K. Yuan, R. Wen, and Z. Li, “Multicontact motion retarget- ing using whole-body optimization of full kinematics and sequential force equilibrium,” Trans. on Mechatronics , vol. 27, no. 5, 2022

    2022

  38. [46]

    Collaborative bimanual manipulation using optimal motion adaptation and interaction control,

    R. Wen, Q. Rouxel, M. Mistry, Z. Li, and C. Tiseo, “Collaborative bimanual manipulation using optimal motion adaptation and interaction control,” IEEE RAM , 2023

    2023

  39. [47]

    Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures,

    L. Espeholt et al. , “Impala: Scalable distributed deep-rl with importance weighted actor-learner architectures,” in ICML, 2018

    2018

  40. [48]

    On the continuity of rotation representations in neural networks,

    Y . Zhou, C. Barnes, J. Lu, J. Yang, and H. Li, “On the continuity of rotation representations in neural networks,” in IEEE/CVF CVPR, 2019

    2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.