Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Group Q-score reweighted matching turns a pretrained navigation diffusion policy into a cross-embodiment agent that recovers from dead ends and hard layouts.

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 · grok-4.5

2026-07-31 03:40 UTC pith:CVRBUXC4

load-bearing objection Solid systems recipe for RL post-training of navigation diffusion policies: group-normalized reweighting plus structured self-perturbation, with real gains and honest limits. the 3 major comments →

arxiv 2607.28560 v1 pith:CVRBUXC4 submitted 2026-07-30 cs.RO

X-NavDP: Generalizing Navigation Diffusion Policy to Novel Behavior and Embodiments with Group Q-score Reweighted Matching

Tianyu Yang , Yiming Zeng , Wenzhe Cai , Yuqiang Yang , Jiaqi Peng , Hui Cheng , Jiangmiao Pang , Tai Wang This is my paper
classification cs.RO
keywords visual navigationdiffusion policyreinforcement learningcross-embodimentscore matchingGQRMtrajectory perturbation
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.

Pretrained navigation diffusion policies copy oracle trajectories for one nominal robot, so they stall on dead ends, long obstacles, and new body types when they only see local onboard images. This paper argues that a short online RL post-training stage can fix that without throwing away the pretrained prior, if two design choices are made carefully. First, exploration mixes the policy’s own goal-conditioned paths with its goal-agnostic paths and light sign flips, so candidates stay smooth and feasible while covering reverse, lateral, and detour moves. Second, Q-values are normalized inside each same-state candidate group and used as weights for score matching, so hard low-return states still teach the actor which local option is better. The resulting policy, X-NavDP, trained across wheeled, quadruped, and humanoid robots in parallel simulation, raises held-out success from 61% to 84% in simulation and from about 10% to 65% in hard real lab, hall, and office trials with no extra real fine-tuning.

Core claim

Stable RL post-training of a navigation diffusion policy is achievable by self-bootstrapped trajectory perturbation plus within-group Q-score reweighted score matching (GQRM), which reallocates probability toward higher-value same-state candidates without backpropagating through the full denoising likelihood. With light embodiment conditioning and chunk-level rewards, one shared policy then generalizes across robot bodies and learns recovery and detour behaviors that pure imitation never showed.

What carries the argument

GQRM (Group Q-score Reweighted Matching): for each state, sample a group of perturbed trajectory chunks from the current policy, normalize their Q-values inside that group, and use the exponential of those normalized scores as weights in a denoising score-matching update so the actor improves locally even when absolute returns are low.

Load-bearing premise

Useful recovery and detour moves already appear among the perturbed candidates drawn from the pretrained policy; group reweighting can only promote what that candidate set contains, not invent actions outside it.

What would settle it

Remove both goal-agnostic mixing and reverse sign flips during exploration (keeping RL and group reweighting): if success collapses as in the paper’s ablation, the claim that structured self-bootstrapped candidates are what unlock hard-state learning is supported; if performance still rises, that premise is false.

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

If this is right

  • Imitation-only navigation diffusion policies can be upgraded with hours of parallel sim RL rather than larger oracle datasets alone.
  • Same-state group value normalization is a practical fix when minibatch Q-weighting lets easy states dominate hard recovery states.
  • One shared diffusion navigator plus light FiLM robot embeddings can serve wheeled, quadruped, and humanoid bases when low-level locomotion controllers already exist.
  • Real deployment of the sim-only post-trained weights can jump hard-layout success without robot-specific real fine-tuning.
  • Failure modes that remain (glass, hollow obstacles, long memory around walls, bulky humanoid clearance) point to perception and memory upgrades rather than a different RL objective.

Where Pith is reading between the lines

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

  • The same within-group reweighting idea may transfer to other long-horizon diffusion policies (manipulation, driving) where absolute returns vary wildly across states.
  • If candidate groups systematically miss a needed recovery mode, pairing GQRM with a small learned skill library or stronger memory could close the remaining wall-detour and narrow-passage gaps.
  • Chunk-level MPC tracking makes the method depend on a competent low-level stack; new morphologies without that stack would need more than the FiLM embedding alone.

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 GQRM (Group Q-score Reweighted Matching), an online RL post-training framework for pretrained navigation diffusion policies. It combines (i) self-bootstrapped trajectory perturbation that mixes goal-conditioned and goal-agnostic samples with signed flips (Eqs. 7–8) and (ii) same-state group-normalized Q-score weights for reweighted score matching (Eqs. 9–11), extending DPMD. Embodiment FiLM modulation and RTC-style temporal guidance support cross-robot closed-loop execution. On 40 held-out IsaacLab scenes across wheeled, quadruped, and humanoid embodiments, X-NavDP raises overall SR/SPL from 61.20%/58.95% (NavDP) to 84.28%/77.19% (Table 1); real-world zero-shot hard cases improve from roughly 10% to 65% average success (Table 2). Ablations compare SFT/DPPO/DSRL/DPMD-original, exploration variants, embodiment conditioning, and RTC (Tables 3–7, Fig. 4).

Significance. If the reported gains hold under fuller statistical reporting, this is a strong and timely contribution to visual navigation and diffusion-policy RL. It targets a genuine gap: imitation from oracle global planners yields brittle local reactive behavior and weak cross-embodiment transfer. The technical path—structured self-bootstrapped candidates plus within-group Q reweighting rather than likelihood-ratio gradients through the reverse chain—is well motivated, and Appendix A cleanly derives GQRM from DPMD. Strengths include multi-embodiment held-out simulation, method ablations that isolate exploration collapse and group normalization, qualitative recovery/detour behaviors, zero-shot real deployment on three platforms, public code/model, and an honest limitations section. The work is primarily empirical systems research; significance rests on reproducibility of the large SR gains and on the claim that GQRM plus structured perturbation is a practical recipe for diffusion navigation post-training.

major comments (3)
  1. [§5.2, Tables 1–3; Appendix B] Tables 1–3 and the real-world Table 2 report only point estimates (SR/SPL) with no seeds, standard errors, or confidence intervals, despite a hierarchical stack (diffusion policy + MPC + embodiment locomotion) and a heavily shaped multi-term reward (Appendix B.1, Table 4). For an RL post-training claim of +23 pp overall SR in simulation and ~10%→65% in real hard cases, multi-seed means and variability are load-bearing. Please add at least 3 independent training seeds for the main X-NavDP vs NavDP/DPMD-original comparisons on the 24- or 56-scene protocol, and clarify how many evaluation episodes underlie each sim cell.
  2. [§4.2–4.3, Eqs. 7–11; Fig. 4; Table 6; Abstract] The central mechanism only reallocates mass among same-state candidates from self-bootstrapped perturbation (Eqs. 7–11); it cannot invent recovery/detour actions absent from the group. Fig. 4 / Table 6 make this explicit: removing both no-goal perturbation and reverse collapses overall SR/SPL to ~6%/5%. The abstract and Fig. 1 framing (“enables new skills such as failure recovery”) should be tightened to state that new behaviors emerge only when the perturbed pretrained prior already covers them, and GQRM amplifies locally better candidates. A short discussion of failure modes when all group members are non-viable would strengthen the claim’s scope.
  3. [§5.1–5.2, Table 2; Abstract] Real-world evaluation uses N=10 trials per cell (lab/hall/office × three embodiments) with no uncertainty and a coarse “hard cases ~10%→65%” aggregate in the abstract. Given embodiment-specific perception failures noted in §5.3 (glass, hollow obstacles, G1 bulk/turning), please report per-cell counts with binomial CIs or equivalent, define the “hard case” selection protocol relative to NavDP failures, and state whether any real-world hyperparameter (e.g., η_guide, speed limits) was tuned on the test platforms.
minor comments (5)
  1. [§4.2–4.3] Notation collision: λ is used both as the signed mixing coefficient in Eq. 7 and as the PMD/GQRM temperature in Eqs. 3–4, 6, and 11. Rename one (e.g., λ_mix vs λ_temp) for clarity.
  2. [Appendix B.2 vs §4.3] In Appendix B.2 the text says candidate values are “normalized by the value dispersion across the minibatch,” which appears to contradict the same-state group normalization emphasized in §4.3. Please reconcile the implementation description with Eq. 9–10.
  3. [Table 1] Table 1 leaves several baseline cells as “–” (NavOL, SIDP, NavDP-RL) on embodiments/scenes where X-NavDP is reported. Briefly justify missing entries (unavailable code, embodiment-specific methods) so the SOTA comparison is not over-read.
  4. [Figure 3] Fig. 3 qualitative trajectories are compelling but would benefit from a short caption note on how scenes/goals were selected (cherry-picked hard cases vs random held-out failures).
  5. [§5.1; front matter] Minor typos/consistency: arXiv ID line and project URL formatting; “wheeled, quadruped, and humanoid” embodiment naming is consistent, but Dingo vs Turtlebot sim-to-real wheeled mismatch could be stated once in §5.1.

Circularity Check

0 steps flagged

No significant circularity: empirical RL post-training paper; GQRM modifies external DPMD weights and claims are held-out/real-robot measurements, not tautologies.

full rationale

X-NavDP’s central claims are empirical performance gains (Tables 1–2: SR/SPL on held-out IsaacLab scenes and real lab/hall/office trials) from an RL post-training recipe, not first-principles predictions forced by fitted inputs. Appendix A derives LGQRM from the external DPMD reweighted score-matching objective by substituting same-state group-normalized advantages eQG for raw Q; the policy-improvement target remains the KL-regularized mirror-descent distribution πMD, so the update is a design choice plus Monte Carlo estimation, not X defined as Y. Self-bootstrapped perturbation (Eq. 7–8) and embodiment FiLM are architectural priors validated by ablations (Fig. 4, Tables 3, 5–7), including collapse when exploration is removed—evidence of contingency, not circularity. Citation of NavDP is backbone reuse with independent evaluation against external baselines (iPlanner, ViPlanner, DPPO, DSRL, DPMD), not a load-bearing uniqueness theorem from overlapping authors. No fitted parameter is renamed a prediction; no ansatz is smuggled in as a theorem. Score 0 is appropriate.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on standard diffusion/RL math, a stack of engineering assumptions (privileged pretraining, hierarchical control, simulator fidelity), and many hand-chosen hyperparameters in exploration, reweighting, and reward. No new physical entities; the invented objects are algorithmic modules whose only evidence is the reported tasks.

free parameters (6)
  • trajectory mix coefficient λ and Bernoulli flip probability ϵ
    Control how goal-agnostic paths and axis sign flips perturb rollouts (Eq. 7–8); chosen for exploration, not derived.
  • group advantage scale c, clip h, temperature λ, ε floor = clip to [-3,3]; top-5 positive; other values not uniquely fixed in main text
    Define eQG and exp weights in GQRM (Eq. 10–11); appendix uses clip [-3,3], top-k positive advantages, learned temperature.
  • multi-term reward weights and thresholds = see Appendix B.1 Table 4
    Living penalty, progress scale 0.075, arrival 2.0, collision 2.0, immobility 0.2, G1 fall -0.4, force thresholds, A* clearance costs (Table 4 / Eq. 37)—hand-designed shaping that steers what RL reinforces.
  • macro-step horizon and embodiment reward normalizers κ_e = 3 s; κ_e = 1 / 2.5 / 2.5
    3 s chunk tracking and κ_e ∈ {1, 2.5, 2.5} rescale returns across robots (Eq. 38).
  • RTC guidance strength η_guide and prefix weights w_t
    Inference-time temporal consistency (Eq. 14–15); ablated but still a free deployment knob.
  • optimizer/schedule knobs (lr 2e-5→1e-5, γ=0.99, τ=0.005, candidate N=64, etc.) = Appendix B.2
    Standard but numerous training choices that affect whether the reported lift is recovered.
axioms (5)
  • standard math Denoising score matching recovers the marginal score of the noised action distribution (Tweedie / DSM equivalence).
    Appendix A.1 uses this to justify regressing ϵ/score with reweighted samples.
  • domain assumption Policy mirror descent closed form π_MD ∝ π_old exp(Q/λ) is the right improvement target for the actor.
    Sec. 3.2–4.3 import PMD/DPMD; navigation success is not guaranteed to optimize under the shaped Q used in training.
  • domain assumption Waypoint-chunk actions tracked by MPC plus embodiment-specific locomotion faithfully realize intended navigation behavior.
    Sec. 4.1 hierarchical stack; limitations note new morphologies need pretrained low-level controllers.
  • domain assumption IsaacLab/GRScenes rollouts and onboard RGB-D depth are close enough to real hard cases for zero-shot transfer claims.
    Sim-to-real tables assume perception and dynamics gap is bridgeable without real fine-tuning.
  • ad hoc to paper Same-state group normalization of Q yields a better local improvement direction than batch-global Q normalization in hard states.
    Central methodological bet of GQRM (Sec. 4.3); supported by DPMD-original ablation but not proved generally.
invented entities (3)
  • GQRM (Group Q-score Reweighted Matching) objective no independent evidence
    purpose: Replace raw/batch Q weights in DPMD with within-group normalized exponential weights for diffusion actor updates.
    Algorithmic construct defined in Eq. 9–11 / Appendix A.2; evidence is task performance, not an external physical prediction.
  • Self-bootstrapped mixed trajectory τ_mixed no independent evidence
    purpose: Generate structured exploration candidates from pointgoal and nogoal branches with signed flips.
    Defined in Eq. 7–8; validated by ablation collapse when removed, not by independent theory.
  • Embodiment FiLM + action-token shift module no independent evidence
    purpose: Condition one shared diffusion decoder on robot identity for cross-embodiment training.
    Architectural add-on in Sec. 4.4; compared to soft-prompt concat in Table 5.

pith-pipeline@v1.2.0-daily-grok45 · 24437 in / 4028 out tokens · 87879 ms · 2026-07-31T03:40:35.670198+00:00 · methodology

0 comments
read the original abstract

Pretraining navigation diffusion policies rely on large-scale expert demonstrations. These data are typically generated by a fully-informed oracle planner suited to a single nominal robot. This limits the policy's generalization to diverse embodiments and challenging scenarios (e.g., escaping dead ends or detouring long obstacles) that demand diverse local reactive behaviors with only onboard local observations. Post-training the policy with reinforcement learning (RL) offers a principled remedy. However, previous RL for diffusion approaches lead to only marginal improvements. This is because the intractable likelihood of diffusion policies renders policy gradients unstable in addition to inefficient policy exploration. To address these challenges, we propose a data-efficient diffusion RL post-training framework - GQRM (Group Q-score Reweighted Matching). Our framework introduces two complementary designs: (i) a self-bootstrapped exploration strategy with behavior perturbation that preserves the pretrained policy prior, and (ii) a group Q-score normalization mechanism that computes per-trajectory values on each state for efficient reweighted score matching. By conducting distributed online RL training across heterogeneous embodiments, the resulting fine-tuned policy, X-NavDP, achieves state-of-the-art cross-embodiment visual navigation performance, improving the overall success rate from 61.20% to 84.28% in simulation and 10% to 65% in real-world hard cases. The code and model are publicly available at https://yty-sky.github.io/x-navdp-project-page.

Figures

Figures reproduced from arXiv: 2607.28560 by Hui Cheng, Jiangmiao Pang, Jiaqi Peng, Tai Wang, Tianyu Yang, Wenzhe Cai, Yiming Zeng, Yuqiang Yang.

Figure 1
Figure 1. Figure 1: We develop an RL post-training framework that improves pretrained navigation diffusion [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the X-NavDP pipeline. (a) We adopt the NavDP backbone and inject em [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative trajectory comparison. X-NavDP produces recovery, detour, and safer behav [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on exploration policies. The results demonstrate that no-goal perturbation [PITH_FULL_IMAGE:figures/full_fig_p008_4.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

43 extracted references · 15 linked inside Pith

  1. [1]

    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. In2026 International Conference on Robotics and Automation (ICRA), 2026

    2026

  2. [2]

    J. Peng, W. Cai, Y . Yang, T. Wang, Y . Shen, and J. Pang. Logoplanner: Localization grounded navigation policy with metric-aware visual geometry. In2026 International Conference on Robotics and Automation (ICRA), 2026

    2026

  3. [3]

    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), pages 11994–12001, 2025

    2025

  4. [4]

    X. Wei, C. Gu, and L. Zhang. Navol: Navigation policy with online imitation learning.arXiv preprint arXiv:2605.11762, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Z. Xie, Z. Liu, Z. Peng, W. Wu, and B. Zhou. Vid2sim: Realistic and interactive simulation from video for urban navigation. InProceedings of the Computer Vision and Pattern Recogni- tion Conference, pages 1581–1591, 2025

    2025

  6. [6]

    M. Wei, C. Wan, J. Peng, X. Yu, Y . Yang, D. Feng, W. Cai, C. Zhu, T. Wang, J. Pang, and X. Liu. Ground slow, move fast: A dual-system foundation model for generalizable vision- language navigation. InThe Fourteenth International Conference on Learning Representa- tions, 2026

    2026

  7. [7]

    H. Ren, Y . Zeng, Z. Bi, Z. Wan, J. Huang, and H. Cheng. Prior does matter: Visual navigation via denoising diffusion bridge models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12100–12110, 2025

    2025

  8. [8]

    Intelligence, A

    P. Intelligence, A. Amin, R. Aniceto, A. Balakrishna, K. Black, K. Conley, G. Connors, J. Darpinian, K. Dhabalia, J. DiCarlo, et al. Pi ∗ 0.6: a vla that learns from experience.arXiv preprint arXiv:2511.14759, 2025. 9

    Pith/arXiv arXiv 2025

  9. [9]

    C. Yu, Y . Wang, Z. Guo, H. Lin, S. Xu, H. Zang, Q. Zhang, Y . Wu, C. Zhu, J. Hu, et al. Rlinf: Flexible and efficient large-scale reinforcement learning via macro-to-micro flow transforma- tion.arXiv preprint arXiv:2509.15965, 2025

    arXiv 2025

  10. [10]

    Sheng, R

    J. Sheng, R. Bai, K. Xu, R. Liu, J. Chen, S. Yuan, W.-Y . Yau, and L. Xie. Beyond imitation: Reinforcement learning fine-tuning for adaptive diffusion navigation policies.arXiv preprint arXiv:2603.12868, 2026

    arXiv 2026

  11. [11]

    C. Xu, J. T. Springenberg, M. Equi, A. Amin, A. Esmail, S. Levine, and L. Ke. Rl token: Bootstrapping online rl with vision-language-action models.arXiv preprint arXiv:2604.23073, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Y . Chen, S. Tian, S. Liu, Y . Zhou, H. Li, and D. Zhao. Conrft: A reinforced fine-tuning method for vla models via consistency policy.arXiv preprint arXiv:2502.05450, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Black, M

    K. Black, M. Janner, Y . Du, I. Kostrikov, and S. Levine. Training diffusion models with reinforcement learning. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  14. [14]

    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. InInternational Conference on Learning Representations, pages 77288–77329, 2025

    2025

  15. [15]

    Wagenmaker, M

    A. Wagenmaker, M. Nakamoto, Y . Zhang, S. Park, W. Yagoub, A. Nagabandi, A. Gupta, and S. Levine. Steering your diffusion policy with latent space reinforcement learning.arXiv preprint arXiv:2506.15799, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    H. Ma, T. Chen, K. Wang, N. Li, and B. Dai. Efficient online reinforcement learning for diffusion policy. InProceedings of the International Conference on Machine Learning (ICML),

  17. [17]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  18. [18]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  19. [19]

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems (RSS), 2023

    2023

  20. [20]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations.arXiv preprint arXiv:2403.03954, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    T.-W. Ke, N. Gkanatsios, and K. Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations.arXiv preprint arXiv:2402.10885, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Sridhar, D

    A. Sridhar, D. Shah, C. Glossop, and S. Levine. NoMaD: Goal masked diffusion policies for navigation and exploration. InInternational Conference on Robotics and Automation (ICRA), 2024

    2024

  24. [24]

    H. Li, Y . Zuo, J. Yu, Y . Zhang, Z. Yang, K. Zhang, X. Zhu, Y . Zhang, T. Chen, G. Cui, et al. Simplevla-rl: Scaling vla training via reinforcement learning.arXiv preprint arXiv:2509.09674, 2025. 10

    Pith/arXiv arXiv 2025

  25. [25]

    Shribak, C.-X

    D. Shribak, C.-X. Gao, Y . Li, C. Xiao, and B. Dai. Diffusion spectral representation for rein- forcement learning.arXiv preprint, 2024

    2024

  26. [26]

    Schulman, S

    J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz. Trust region policy optimization. InInternational Conference on Machine Learning (ICML), 2015

    2015

  27. [27]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. InarXiv preprint arXiv:1707.06347, 2017

    Pith/arXiv arXiv 2017

  28. [28]

    Tomar, L

    M. Tomar, L. Shani, Y . Efroni, and M. Ghavamzadeh. Mirror descent policy optimization. In International Conference on Learning Representations (ICLR), 2022

    2022

  29. [29]

    G. Lan. Policy mirror descent for regularized reinforcement learning: A generalized frame- work with linear convergence.SIAM Journal on Optimization, 33(2):1061–1091, 2023

    2023

  30. [30]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. InConference on robot learning, pages 91–100. PMLR, 2022

    2022

  31. [31]

    Mayor, J

    W. Mayor, J. Obando-Ceron, A. Courville, and P. S. Castro. The impact of on-policy parallelized data collection on deep reinforcement learning networks.arXiv preprint arXiv:2506.03404, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Mittal, P

    M. Mittal, P. Roth, J. Tigue, A. Richard, O. Zhang, P. Du, A. Serrano-Munoz, X. Yao, R. Zurbr ¨ugg, N. Rudin, et al. Isaac lab: A gpu-accelerated simulation framework for multi- modal robot learning.arXiv preprint arXiv:2511.04831, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Perez, F

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville. FiLM: Visual reasoning with a general conditioning layer. InAAAI Conference on Artificial Intelligence, 2018

    2018

  34. [34]

    Black, M

    K. Black, M. Galliker, and S. Levine. Real-time execution of action chunking flow policies. Advances in Neural Information Processing Systems, 38:33383–33407, 2026

    2026

  35. [35]

    Mittal, C

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg. Orbit: A unified simulation framework for interactive robot learning environments. InIEEE Robotics and Automation Letters, 2023

    2023

  36. [36]

    Zhong, P

    W. Zhong, P. Cao, Y . Jin, L. Li, W. Cai, J. Lin, Z. Lyu, T. Wang, B. Dai, X. Xu, and J. Pang. In- ternscenes: A large-scale interactive indoor scene dataset with realistic layouts. InConference on Neural Information Processing Systems (NeurIPS), 2025

    2025

  37. [37]

    Contributors

    I. Contributors. InternNav: InternRobotics’ open platform for building generalized navigation foundation models.https://github.com/InternRobotics/InternNav, 2025

    2025

  38. [38]

    Anderson, A

    P. Anderson, A. X. 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. InarXiv preprint arXiv:1807.06757, 2018

    Pith/arXiv arXiv 2018

  39. [39]

    F. Yang, C. Wang, C. Cadena, and M. Hutter. iplanner: Imperative path planning.arXiv preprint arXiv:2302.11434, 2023

    arXiv 2023

  40. [40]

    P. Roth, J. Nubert, F. Yang, M. Mittal, and M. Hutter. Viplanner: Visual semantic impera- tive learning for local navigation. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5243–5249. IEEE, 2024

    2024

  41. [41]

    Zhang, J

    R. Zhang, J. Hou, C. Cheng, Q. Chen, T. Wang, and W. Zhao. Self-imitated diffusion policy for efficient and robust visual navigation.arXiv preprint arXiv:2601.22965, 2026. 11

    Pith/arXiv arXiv 2026

  42. [42]

    J. Wang, M. Chen, S. Zhang, N. Karaev, J. Sch¨onberger, P. Labatut, P. Bojanowski, D. Novotny, A. Vedaldi, and C. Rupprecht. Vggt-omega.arXiv preprint arXiv:2605.15195, 2026

    Pith/arXiv arXiv 2026

  43. [43]

    w/o extrapolation

    B. Efron. Tweedie’s formula and selection bias.Journal of the American Statistical Associa- tion, 106(496):1602–1614, 2011. 12 A GroupQ-score Reweighted Matching Derivation We derive the GQRM actor objective by building on the reweighted score-matching framework of DPMD. Starting from the DPMD objective, GQRM replaces the raw Q-score weight with a same-st...

    2011