Pith. sign in

REVIEW 3 major objections 5 minor 33 references

Under a matched training budget, pairing few-step MeanFlow generation with critic-guided refinement and PPO fine-tuning lifts visual-navigation success to 74.7% and cuts per-plan latency from 85 ms to 60 ms.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 01:28 UTC pith:NIOZ3L6I

load-bearing objection The paper is a competent systems integration of MeanFlow, critic guidance, and PPO for navigation, but the headline gain over NavDP is not attributable to the proposed components because there is no NavDP+RL control. the 3 major comments →

arxiv 2607.14643 v1 pith:NIOZ3L6I submitted 2026-07-16 cs.RO

NavCMPO: Critic-Guided MeanFlow Policy Optimization for Adaptive Navigation

Junjie An , Yi Wu , Xiao Liu , Yiqun Zhou , Yuechen Wu , Xiaoqing Guan , You Wang , Guang Li This is my paper
classification cs.RO
keywords visual navigationMeanFlowflow matchingdiffusion policyPPO fine-tuningcritic-guided refinementobstacle proximity predictionsim-to-real transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes NavCMPO, a two-stage framework for mapless visual navigation built on three ideas: generate trajectories with a five-step MeanFlow policy instead of ten-step diffusion denoising; compensate for the obstacle-avoidance degradation that few-step generation causes by using gradients from a critic trained with obstacle-point-cloud supervision (CGTR); and fine-tune the policy with PPO plus behavior-cloning regularization so it can exceed expert-demonstration quality. The authors report that under matched compute on the InternVLA-N1 benchmark, NavCMPO reaches 74.7% average success rate, 6.4 percentage points above a retrained NavDP baseline, while reducing inference latency from 85 ms to 60 ms. Real-world trials on a Unitree Go2 quadruped show 66.7% success versus 56.7% for NavDP. The argument is that each component is load-bearing, but the authors explicitly note that the baseline did not receive the same RL fine-tuning, so the incremental credit is not fully isolated.

Core claim

NavCMPO is a local planner that maps RGB-D history plus a goal waypoint to a trajectory. In Stage 1, a MeanFlow policy head predicts a velocity field transporting Gaussian noise to expert trajectories; an Obstacle Proximity Prediction (OPP) auxiliary task forces the visual encoder to output per-sector distances to the nearest obstacle (12 uniform angular sectors, SmoothL1 loss); and a critic head, sharing the transformer decoder, is trained with mean-squared error against a hand-defined score that penalizes waypoints closer than dsafe to obstacle points and rewards goal progress. Because five-step MeanFlow loses spatial reasoning compared to ten-step diffusion, CGTR takes one clipped gradien

What carries the argument

The load-bearing object is the critic-coupled few-step MeanFlow policy: a flow-matching head that generates trajectory waypoints in T=5 integration steps, paired with a critic head trained on obstacle-point-cloud-derived scores (safety penalty plus goal-progress bonus) and a single-step CGTR correction (z_tau <- z_tau + eta * Clip(grad_z C, delta)) applied at the penultimate step. The OPP auxiliary task (12-sector nearest-obstacle distance prediction) and the Stage-2 PPO objective with BC regularization (L_BC = E ||tau_ref - tau_theta||^2) are supporting mechanisms that make the core work.

Load-bearing premise

The comparison benchmark is a 'retrained NavDP baseline' that does not receive the paper's Stage-2 RL fine-tuning, so the reported advantage could be entirely due to generic fine-tuning, not the MeanFlow/CGTR components.

What would settle it

Apply exactly the paper's Stage-2 procedure (same rewards, BC anchor, critic update, same compute) to NavDP's ten-step diffusion policy. If NavDP+PPO reaches within one percentage point of NavCMPO's 74.7%, then the MeanFlow and CGTR components are not the source of the headline gain.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If NavCMPO is right, diffusion-based navigation policies do not need ten denoising steps; five MeanFlow steps plus one critic-gradient correction recover the safety of ten-step denoising at 60 ms per plan.
  • RL fine-tuning of a flow-based policy with a frozen BC anchor can surpass the demonstration ceiling without catastrophic forgetting, raising success by 7.5 points over the behavior-cloned version in the paper's ablations.
  • A critic trained offline with geometric supervision can act as a lightweight, online-free safety layer for few-step generative planners, an alternative to explicit cost-map optimization at inference.
  • The 12-sector obstacle-proximity pretraining task transfers across embodiments: real-world success on the Unitree Go2 rises to 66.7% from the 56.7% of NavDP, concentrated in hard and outdoor scenes.

Where Pith is reading between the lines

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

  • The paper's matched-compute comparison does not include a control in which the same PPO fine-tuning routine is applied to NavDP, so the 6.4-point gain is not evidence that MeanFlow plus CGTR is superior to the diffusion backbone under RL fine-tuning; a direct control would tell.
  • A natural extension is to measure the contribution of CGTR alone by applying the same Stage-2 fine-tuning to a five-step MeanFlow policy without the critic correction; the paper's ablation (w/o CGTR, 60.0% vs 73.5%) already suggests CGTR is important, but that ablation is on the pre-trained policy, not the fully fine-tuned one.
  • The 60 ms figure is still five full transformer decoder passes; mapping the quality-latency frontier for one- and three-step configs with CGTR (the paper tests one-step but only for the pre-trained policy) could tell whether the robot can plan at 30 ms with acceptable safety.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes NavCMPO, a two-stage navigation policy that combines few-step MeanFlow trajectory generation with an obstacle-proximity-prediction (OPP) auxiliary task, a critic-guided trajectory refinement (CGTR) mechanism, and PPO fine-tuning with behavior-cloning regularization. In Stage 1, a MeanFlow policy is pre-trained with a critic head supervised by an explicit geometric score based on obstacle point-cloud distances and goal progress. CGTR refines the intermediate flow state using critic gradients. In Stage 2, the policy is fine-tuned with PPO while the critic is updated and a frozen BC reference regularizes the policy. Experiments on the InternVLA-N1 benchmark (IsaacSim, Dingo robot) report a 74.7% average success rate versus 68.3% for a retrained NavDP baseline, with inference latency reduced from 85 ms to 60 ms, and a real-world unitree Go2 study reports 66.7% versus 56.7% success. The paper concludes by acknowledging that it does not compare against applying the same RL fine-tuning procedure to NavDP.

Significance. The paper addresses a practically relevant problem: diffusion-based navigation policies are accurate but slow, and few-step generative models degrade obstacle avoidance. The proposed recipe -- few-step MeanFlow with a critic trained on explicit geometric scores, plus PPO fine-tuning -- is a sensible integration of existing techniques, and the explicit geometric supervision of the critic is a transparent design that avoids circularity. If the reported gains are robust, the framework would be a useful step toward low-latency, embodiment-adaptive navigation. The matched-compute comparison with a retrained NavDP baseline is a strength, as is the inclusion of a real-robot transfer study. However, the current evidence does not isolate the contribution of the proposed critic-guided MeanFlow components from generic RL fine-tuning of a strong BC policy, and the lack of repeated trials and a clear validation protocol substantially weakens confidence in the quantitative claims. The paper is therefore promising but not yet conclusive.

major comments (3)
  1. The central claim that NavCMPO outperforms NavDP by 6.4 pp is not supported as stated because there is no control applying the same PPO fine-tuning procedure to NavDP. Decomposing Table I, NavCMP without RL fine-tuning is 69.2% versus 68.3% for NavDP (a 0.9 pp gap), while NavCMPO is 74.7% -- a 5.5 pp jump attributed to RL fine-tuning. Since prior work (DPPO, DMPO) already shows that PPO fine-tuning improves diffusion/flow policies beyond BC, the headline result is equally consistent with 'generic RL fine-tuning of any strong BC policy helps'. The Conclusion explicitly admits this limitation. The paper should add a NavDP+PPO control (ideally with the same BC loss and critic-update choices) and report the resulting SR/SPL. Without this, the load-bearing attribution to the proposed critic-guided MeanFlow architecture is not established.
  2. All empirical results are single-run point estimates with no error bars, confidence intervals, or multiple seeds. This is particularly problematic for the small absolute differences: e.g., the real-world comparison in Table III is 20/30 vs. 17/30 (NavCMPO vs. NavDP), a difference of 3 successes that could easily be sampling noise. In addition, the default hyperparameters (eta=0.4, delta=1.0, one CGTR step, five denoising steps) are selected using sensitivity analyses (Figs. 4-5) that appear to be run on the same evaluation scenes used for the final reported numbers, with no described train/validation split. This creates an optimistic bias in the benchmark results. The manuscript should report means and standard deviations over at least 5 seeds and provide a clear separation between validation scenes used for hyperparameter selection and held-out test scenes.
  3. The ablation and comparison tables use different evaluation protocols, making direct cross-table claims misleading. Table II reports an average of 73.5% for NavCMPO, while Table I reports 74.7%, and the text compares FlowNav's average 51.0% (from the 100-scene Easy/Hard protocol) with MP1's 59.1% (from the Cluttered-Hard/Home/Commercial protocol) without noting the mismatch. Also, the text mentions Cluttered-Easy as a test environment, but Table I has no Cluttered-Easy column. The authors should clarify the exact evaluation protocol used for each table and avoid direct numerical comparisons across different scene sets.
minor comments (5)
  1. Equation (2) appears mathematically malformed: "d/dt u(z_t,r,t) = v(z_t,t) ∂_z u + ∂_t u" is not the standard MeanFlow/flow-matching velocity-field equation. This is likely a typo, but it should be corrected or removed.
  2. The log-probability formula for the CGTR-modified step is unclear: writing log N(z_{τ+1}+C | μ_k, σ²I) suggests that the CGTR correction is added to the observation rather than subtracted from the mean. The paper should define the exact action parameterization and state how the deterministic CGTR transform is accounted for in the PPO probability ratio.
  3. The baseline description calls ViPlanner a "visual semantic instruction planner," but ViPlanner is a visual semantic imperative learning method. Please correct the description.
  4. The left panel shows success rate versus denoising steps, but the text says CGTR adds "minimal overhead (10ms)" while Fig. 3 right appears to show no latency difference between with/without CGTR. Please clarify whether the latency numbers include CGTR and how the 10 ms was measured.
  5. The real-world results are based on 10 trials per condition. A sentence acknowledging the small sample size and lack of statistical significance testing would help calibrate the claim of "effective sim-to-real transfer."

Circularity Check

0 steps flagged

No significant circularity: the critic is a supervised surrogate of an explicit geometric score, the policy is optimized against environment rewards, and the admitted missing NavDP+RL control is a comparison confound, not a circular step.

full rationale

The paper's central result is an empirical benchmark comparison, not a derivation from fitted values. The critic score s(τ) (Eq. 5) is computed directly from obstacle-point-cloud distances and goal progress; the critic is a supervised regressor (Eq. 6) trained to predict that external score, and CGTR (Eq. 8) performs gradient ascent on this learned surrogate. This is cost-surrogate optimization, not a definitional or self-predicting loop: the target being optimized is not the policy's own output. Stage-1 losses (Eq. 7) combine MeanFlow regression (Eq. 3), OPP with point-cloud-derived distance labels (Eq. 1), and critic regression. Stage-2 PPO (Eq. 14) is driven by environment rewards (Eqs. 12–13), with BC regularization (Eq. 15) anchored to a frozen reference policy. The admitted limitation in Section V — “it does not directly compare against applying the same RL fine-tuning procedure to NavDP” — weakens attribution of the 6.4pp gain to the proposed components versus generic RL fine-tuning, but this is a control/comparison confound, not a circular derivation. Hyperparameter choices for η and δ (Figs. 4–5) are model selection on an evaluation set and could cause optimistic bias, but they are not fitted values renamed as predictions. External references (MeanFlow [9], NavDP [8], DPPO [11], DMPO [12]) are used as building blocks; no load-bearing argument reduces to a self-citation chain. No circular step found.

Axiom & Free-Parameter Ledger

10 free parameters · 6 axioms · 0 invented entities

All empirical gains depend on many hand-set hyperparameters (eta, delta, lambda_c, lambda_g, beta's, T, and unreported loss weights), several of which are tuned on the evaluation scenes. No code/data release and no NavDP+RL control make it hard to separate component contributions from generic RL fine-tuning.

free parameters (10)
  • CGTR guidance coefficient eta = 0.4
    Selected by sensitivity analysis on evaluation scenes (Fig. 5); peak SR at eta=0.4.
  • CGTR gradient clip delta = 1.0
    Fixed clipping threshold in Eq. (8); no sensitivity analysis reported.
  • Critic score weights lambda_c, lambda_g = 5.0, 0.1
    Hand-set in Eq. (5) to prioritize obstacle avoidance over goal progress.
  • RL reward weights beta_p, beta_s, beta_g = 5, 50, 20
    Hand-set in Eq. (12) for progress, safety, and goal reward.
  • Denoising/integration steps T = 5
    Chosen by Fig. 3 ablation as the best quality-latency tradeoff.
  • CGTR guidance steps = 1 (penultimate step)
    Chosen by Fig. 4 ablation; multi-step guidance causes over-correction.
  • OPP angular sectors = 12
    Discretization choice for obstacle proximity prediction (Section III-B).
  • Loss weights lambda_MF, lambda_OPP, lambda_Critic, lambda_aux, lambda_BC, lambda_Critic2 = not reported
    Eqs. (7) and (16) define total losses but the loss weights are never specified.
  • Safety distance d_safe = not reported
    Used in critic score Eq. (5) and reward Eq. (13); value not given.
  • PPO noise sigma and clipping epsilon = not reported
    Gaussian action noise in Eq. (9) and clip range in Eq. (14) are not specified.
axioms (6)
  • standard math Linear-interpolated flow matching with L2 velocity regression yields a valid few-step trajectory generator.
    This is the MeanFlow/flow-matching framework assumed from [9] and [10].
  • domain assumption The hand-defined critic score s(tau) in Eq. (5) is a valid proxy for navigation quality (safety and goal progress).
    The entire CGTR mechanism depends on this score; no evidence is given that it correlates with actual task success.
  • domain assumption Retraining NavDP under matched compute is a fair comparison, and the retrained baseline is representative.
    The 6.4pp claim is relative to the retrained NavDP baseline; without this assumption the claim only compares against the authors' own baseline implementation.
  • domain assumption PPO with Gaussian noise around the MeanFlow prediction and BC regularization improves reward without destroying the pre-trained prior.
    This is the standard DPPO/DMPO recipe; no convergence or stability guarantee is provided.
  • domain assumption Simulation training on a Go2 model transfers to the physical Unitree Go2.
    Real-world experiments are only 10 trials per scene with no domain randomization or safety analysis described.
  • domain assumption Obstacle point-cloud supervision and ESDF rewards in simulation are accurate enough for OPP labels and RL rewards.
    OPP and the critic both depend on these geometric sources; errors would propagate into the learned policy.

pith-pipeline@v1.3.0-alltime-deepseek · 11515 in / 14152 out tokens · 139423 ms · 2026-08-02T01:28:33.695461+00:00 · methodology

0 comments
read the original abstract

End-to-end diffusion-based policies have demonstrated strong performance in mapless visual navigation, but their iterative denoising process introduces substantial inference latency, while behavior cloning limits performance to the quality of expert demonstrations. We present NavCMPO, a two-stage adaptive navigation framework that combines few-step MeanFlow trajectory generation, critic-guided refinement, and reinforcement learning fine-tuning. During pre-training, an obstacle proximity prediction task encourages the visual representation to capture obstacle-aware spatial information. To compensate for the degradation in obstacle avoidance caused by few-step generation, Critic-Guided Trajectory Refinement (CGTR) uses gradients from a critic trained with obstacle-point-cloud supervision to refine intermediate trajectories. During adaptation, the MeanFlow policy is fine-tuned using Proximal Policy Optimization with behavior-cloning regularization, while the critic is updated to accommodate embodiment-specific observation changes. Under a matched training budget on the InternVLA-N1 benchmark, NavCMPO achieves an average success rate of 74.7\%, exceeding the retrained NavDP baseline by 6.4 percentage points, while reducing inference latency from 85\,ms to 60\,ms. Experiments on a Unitree Go2 further demonstrate effective sim-to-real transfer.

Figures

Figures reproduced from arXiv: 2607.14643 by Guang Li, Junjie An, Xiao Liu, Xiaoqing Guan, Yiqun Zhou, Yi Wu, You Wang, Yuechen Wu.

Figure 1
Figure 1. Figure 1: Overview of the two-stage NavCMPO framework. In Stage 1, RGB-D observations and the navigation goal are encoded into a conditional [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative trajectory comparison in two representative indoor [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of CGTR steps on trajectory quality. The top row shows [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Left: Success rate versus denoising steps with and without CGTR. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: PPO fine-tuning reward curves under different numbers of trainable [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Real-world Experiments. (a) Indoor scene: the top row shows the robot traversing a cluttered corridor; the bottom row shows the corresponding [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 6 linked inside Pith

  1. [1]

    Probabilistic robotics,

    S. Thrun, “Probabilistic robotics,”Communications of the ACM, vol. 45, no. 3, pp. 52–57, 2002

    2002

  2. [2]

    Cog- nitive mapping and planning for visual navigation,

    S. Gupta, J. Davidson, S. Levine, R. Sukthankar, and J. Malik, “Cog- nitive mapping and planning for visual navigation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017

    2017

  3. [3]

    On evaluation of embodied navigation agents,

    P. Anderson, A. Chang, D. S. Chaplot, A. Dosovitskiy, S. Gupta, V . Koltun, J. Kosecka, J. Malik, R. Mottaghi, M. Savva, and A. R. Zamir, “On evaluation of embodied navigation agents,” 2018. [Online]. Available: https://arxiv.org/abs/1807.06757

    Pith/arXiv arXiv 2018

  4. [4]

    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,” in2017 IEEE International Conference on Robotics and Automation (ICRA), 2017, pp. 3357–3364

    2017

  5. [5]

    Learning to navigate in complex environments,

    P. Mirowski, R. Pascanu, F. Viola, H. Soyer, A. Ballard, A. Banino, M. Denil, R. Goroshin, L. Sifre, K. Kavukcuoglu, D. Kumaran, and R. Hadsell, “Learning to navigate in complex environments,” inInternational Conference on Learning Representations, 2017. [Online]. Available: https://openreview.net/forum?id=SJMGPrcle

    2017

  6. [6]

    Spatially-enhanced recurrent memory for long-range mapless naviga- tion via end-to-end reinforcement learning,

    F. Yang, P. Frivik, D. Hoeller, C. Wang, C. Cadena, and M. Hutter, “Spatially-enhanced recurrent memory for long-range mapless naviga- tion via end-to-end reinforcement learning,”The International Journal of Robotics Research, 2025

    2025

  7. [7]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,” The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2024

    2024

  8. [8]

    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,” CoRR, vol. abs/2505.08712, 2025. [Online]. Available: https: //doi.org/10.48550/arXiv.2505.08712

    doi:10.48550/arxiv.2505.08712 2025

  9. [9]

    Mean flows for one-step generative modeling,

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He, “Mean flows for one-step generative modeling,” 2025. [Online]. Available: https://arxiv.org/abs/2505.13447

    Pith/arXiv arXiv 2025

  10. [10]

    Mp1: Meanflow tames policy learning in 1-step for robotic manipulation,

    J. Sheng, Z. Wang, P. Li, and M. Liu, “Mp1: Meanflow tames policy learning in 1-step for robotic manipulation,”CoRR, vol. abs/2507.10543, July 2025. [Online]. Available: https://doi.org/10. 48550/arXiv.2507.10543

    doi:10.48550/arxiv.2507.10543 2025

  11. [11]

    Diffusion policy policy optimization,

    A. Z. Ren, J. Lidard, L. L. Ankile, A. Simeonov, P. Agrawal, A. Majumdar, B. Burchfiel, H. Dai, and M. Simchowitz, “Diffusion policy policy optimization,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=mEpqHvbD2h

    2025

  12. [12]

    One step is enough: Dispersive meanflow policy optimization,

    G. Zou, H. Wang, H. Wu, Y . Qian, Y . Wang, and W. Li, “One step is enough: Dispersive meanflow policy optimization,” 2026. [Online]. Available: https://arxiv.org/abs/2601.20701

    arXiv 2026

  13. [13]

    π RL: Online rl fine-tuning for flow-based vision-language-action models,

    K. Chen, Z. Liu, T. Zhang, Z. Guo, S. Xu, H. Lin, H. Zang, X. Li, Q. Zhang, Z. Yu, G. Fan, T. Huang, Y . Wang, and C. Yu, “π RL: Online rl fine-tuning for flow-based vision-language-action models,”

  14. [14]

    Planning with diffusion for flexible behavior synthesis,

    M. Janner, Y . Du, J. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” inInternational Conference on Machine Learning, 2022

    2022

  15. [15]

    Two-steps diffusion policy for robotic manipulation via genetic denoising,

    M. Cl ´emente, L. M. Brunswic, R. H. Yang, X. Zhao, Y . H. Khalil, H. LEI, A. Rasouli, and Y . Li, “Two-steps diffusion policy for robotic manipulation via genetic denoising,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id=YY1MPKBHp7

    2025

  16. [16]

    One-step diffusion policy: Fast visuomotor policies via diffusion distillation,

    Z. Wang, Z. Li, A. Mandlekar, Z. Xu, J. Fan, Y . Narang, L. Fan, Y . Zhu, Y . Balaji, M. Zhou, M.-Y . Liu, and Y . Zeng, “One-step diffusion policy: Fast visuomotor policies via diffusion distillation,”CoRR, vol. abs/2410.21257, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2410.21257

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.21257 2024

  17. [17]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inThe Eleventh International Conference on Learning Representations, 2023. [Online]. Available: https://openreview.net/forum?id=PqvMRDCJT9t

    2023

  18. [18]

    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,”CoRR, vol. abs/2412.04987, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2412.04987

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.04987 2024

  19. [19]

    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,” inNeurIPS 2022 Workshop on Score-Based Methods, 2022. [Online]. Available: https://openreview.net/forum?id=gWxpdtQpiYV

    2022

  20. [20]

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

    S. Gode, A. Nayak, D. N. P. Oliveira, M. Krawez, C. Schmid, and W. Burgard, “FlowNav: Combining flow matching and depth priors for efficient navigation,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2025. [Online]. Available: https://arxiv.org/abs/2411.09524

    Pith/arXiv arXiv 2025

  21. [21]

    NaviDiffusor: Cost-guided diffusion model for visual navigation,

    Y . Zeng, H. Ren, S. Wang, J. Huang, and H. Cheng, “NaviDiffusor: Cost-guided diffusion model for visual navigation,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 11 994–12 001. [Online]. Available: https://arxiv.org/abs/2504. 10003

    2025

  22. [22]

    iplanner: Imperative path planning,

    F. Yang, C. Wang, C. Cadena, and M. Hutter, “iplanner: Imperative path planning,” inRobotics: Science and Systems XIX, ser. RSS2023. Robotics: Science and Systems Foundation, Jul. 2023. [Online]. Available: http://dx.doi.org/10.15607/RSS.2023.XIX.064

    doi:10.15607/rss.2023.xix.064 2023

  23. [23]

    Viplanner: Visual semantic imperative learning for local navigation,

    P. Roth, J. Nubert, F. Yang, M. Mittal, and M. Hutter, “Viplanner: Visual semantic imperative learning for local navigation,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 5243–5249

    2024

  24. [24]

    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,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 63–70

    2024

  25. [25]

    Reinflow: Fine-tuning flow matching policy with online reinforcement learning,

    T. Zhang, C. Yu, S. Su, and Y . Wang, “Reinflow: Fine-tuning flow matching policy with online reinforcement learning,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id= ACagRwCCqu

    2025

  26. [26]

    Flow-GRPO: Training flow matching models via online RL,

    J. Liu, G. Liu, J. Liang, Y . Li, J. Liu, X. Wang, P. Wan, D. ZHANG, and W. Ouyang, “Flow-GRPO: Training flow matching models via online RL,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. [Online]. Available: https://openreview.net/forum?id=oCBKGw5HNf

    2025

  27. [27]

    Accelerating reinforcement learning with learned skill priors,

    K. Pertsch, Y . Lee, and J. Lim, “Accelerating reinforcement learning with learned skill priors,” inConference on robot learning. PMLR, 2021, pp. 188–204

    2021

  28. [28]

    Reinforcement learning control with knowledge shaping,

    X. Gao, J. Si, and H. Huang, “Reinforcement learning control with knowledge shaping,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 3, pp. 3156–3167, 2023

    2023

  29. [29]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.- W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. J ´egou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski, “DINOv2: Learning robust visual features witho...

    2024

  30. [30]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,”Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  31. [31]

    Flow matching policy gradients,

    D. McAllister, S. Ge, B. Yi, C. M. Kim, E. Weber, H. Choi, H. Feng, and A. Kanazawa, “Flow matching policy gradients,” 2025. [Online]. Available: https://arxiv.org/abs/2507.21053

    Pith/arXiv arXiv 2025

  32. [32]

    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,” inInternational Conference on Learning Representations, 2020. [Online]. Available: https: //openreview.net/forum?id=H1gX8C4YPr

    2020

  33. [2026]

    Available: https://arxiv.org/abs/2510.25889

    [Online]. Available: https://arxiv.org/abs/2510.25889

    arXiv