Robotics · arXiv:2610.08640 v1 · submitted 6 Oct 2026 · 38 references
RIWANav: Recursive World-Action Models with Self-Improvement for Urban Navigation
The paper says
“Long-horizon urban navigation requires sequential local decisions whose errors can compound over time. Imitation learning (IL) rarely learns from failures, while physical trial-and-error reinforcement learning (RL) is costly.”
more of the abstract
Action-conditioned world models can provide imagined feedback by predicting visual consequences for candidate actions. However, a frozen world model may become less reliable as the policy evolves. In this paper, we introduce RIWANAV, a post-training framework that casts the coupled adaptation of a world model and an action model (policy) as task-specific recursive self-improvement (RSI). Each cycle alternates two updates. The world model evaluates policy actions through imagined outcomes, providing comparative feedback for group-relative policy optimization (GRPO). The improved policy then constructs a grounded self-curriculum, selecting expert-consistent action-video pairs by behavioral novelty and prediction error. The refined world model supplies feedback for the next policy update, closing the recursive self-improvement loop. Experiments show that RIWANAV outperforms training baselines and prior methods, validating the proposed recursive self-improvement loop between the policy and world model. Real-world trials further demonstrate its practical applicability.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 7 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A controlled, arithmetic-consistent study showing recursive world-model adaptation helps urban navigation, but the reward function driving every policy update is never defined, so the causal story doesn't fully close. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 38 references say.
The paper says, at III.B, Eqs. (2)--(3)
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The navigation scoring function R_nav is never defined. Section III.B states only that r_i = R_nav(s, a_i, o_hat_i) 'measures the quality of its imagined consequence'; no formula, no inputs beyond its signature, no ablation, and no validation of R_nav appear anywhere.
more of objection 1
Since Eq. (3) is the objective for every policy update in Tables I--III, and since R_nav is also the only reward signal, the measured deltas cannot currently be attributed to the claimed recursive mechanism rather than to whatever R_nav happens to reward. The authors should give the precise definition of R_nav (e.g., goal-progress in predicted frames, collision/artifact penalties), and provide at least a sanity check that R_nav correlates with real closed-loop success, ideally an ablation that isolates R_nav's dependence on the predicted frames.
The paper says, at IV.A (baselines)
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
It is unclear what reward the policy-only PPO and GRPO baselines optimize. Section IV.A says they 'optimize the same navigation objective without action-conditioned visual feedback,' but R_nav as written in III.B consumes imagined frames o_hat_i from COSMOS. Either these baselines use a different reward (in which case the GRPO vs GRPO-WM contrast in Table I confounds the optimizer with a reward change), or R_nav is computable without world-model outputs (in which case the imagined channel is decorative and Eqs.
more of objection 2
(2)--(3) are not actually driven by predicted consequences). The paper must state explicitly what reward the policy-only baselines receive, and reconcile that with the definition of R_nav.
The paper says, at IV.C (ablation) and IV.B
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The recursive-loop claim requires showing that the closed-loop gain is not an artifact of a self-fulfilling proxy. Because R_nav is itself the shared objective of GRPO, GRPO-WM, and RIWANAV, and because the world model is adapted to the policy's own proposals (Eqs.
more of objection 3
(5)--(7)), the world model can in principle make imagined rewards more favorable without improving real outcomes. The externally measured closed-loop metrics certify the full pipeline but not the imagined-feedback channel. A concrete test that would resolve this is a comparison of RIWANAV against a variant that uses the identical R_nav but with world-model updates replaced by uniform expert replay, holding the R_nav-dependent GRPO signal fixed; the current Table III rows do not isolate R_nav because they change both the world-model data and its effect on rewards. At minimum, the authors should report a correlation between R_nav scores and achieved success on a held-out episode set.
5 minor objections
The paper says, at Title
Pith says: minor objection 1 of 5
The title reads 'RIWANav' while the body consistently uses 'RIWANAV'; the spelling should be unified.
The paper says, at IV.E
Pith says: minor objection 2 of 5
The real-world deployment is purely qualitative (Fig. 5). The authors call it 'trials' but report no quantitative metrics (success rate, number of runs, environment variety); a sentence stating that no quantitative evaluation was performed would make the scope of this section precise.
show 3 more
The paper says, at III.B, Eq. (3)
Pith says: minor objection 3 of 5
The expert-anchor loss L_exp is referenced in Eq. (3) and in the 'Learning from imagination' paragraph but its precise combination of waypoint and directional terms is only described loosely; a one-line definition would be helpful for reproducibility.
The paper says, at IV.C
Pith says: minor objection 4 of 5
The static-expert-replay row lowering SR (80.31 to 78.23) is reported and explained, which is good practice; the explanation would be stronger if accompanied by a hypothesis about why updating on expert-only data hurts (e.g., distribution shift), even a brief one.
The paper says, at IV.A(c)
Pith says: minor objection 5 of 5
The hyperparameters (sigma, q_min, lambda_N, lambda_E, tau_N, tau_E, w_m, tau_m, alpha, beta, lambda_exp) are listed only for a subset; a table reporting all free parameters used in the experiments would aid reproducibility.
What would settle it
Require the authors to specify R_nav's complete computation and the reward variant used by the policy-only PPO/GRPO baselines. Then run one check that settles the mechanism: on a held-out set of states with candidate groups sampled from the trained policy, compute (a) imagined rewards r_i = R_nav(s, a_i, ô^+_i) from the current world model and (b) the same scoring function on recorded ground-truth futures, and track the rank correlation (Kendall's τ) between the two across the 19 recursion rounds of Fig. 3. If τ stays flat or declines while SR rises, improved imagined feedback is not carrying the gains and the recursive-loop claim fails; if τ climbs, the loop genuinely aligns imagination with reality.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 7 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Two things to know. First, the empirical core is real and unusually disciplined: a controlled comparison fixing architecture, initialization, data, and protocol, with every delta I checked arithmetically consistent, and an ablation that decomposes the gains coherently (test-unseen SR: 80.31 frozen, 78.23 with WM update only, 84.83 with policy-induced replay, 88.25 full). Second, the mechanism the paper claims — that the recursive loop, not just world-model feedback, produces the gains — depends on a reward function that is never defined. Section III.B gives ri = R_nav(s, ai, o_hat^+_i) with no formula, no ablation, no validation. The reader is right to make this the central condition.
What's genuinely new: the specific combination of expert-consistency filtering plus novelty- and error-prioritized replay for an action-conditioned video world model inside a GRPO loop. Each ingredient is prior art, but the integration is new, and the external grounding is real — closed-loop success is measured against the actual URBANNAV environment and world-model fidelity against held-out frames, so the headline is not definitional.
Soft spots, in order of weight. R_nav first: every policy update and every baseline in Tables I–III is a difference of training under this unspecified objective, so the deltas cannot be fully attributed to the imagined-feedback mechanism. Second, the policy-only baselines are described as optimizing the same navigation objective without visual feedback, yet R_nav as written consumes imagined frames — the paper never states what reward PPO/GRPO actually use, which is either a confound or a sign the imagined channel is decorative. Third, the self-curriculum hyperparameters (q_min = 0.35, lambda_N = lambda_E = 0.50, the tau scales) are hand-set with no sensitivity analysis, and no code or models are released, so the recipe is not reproducible from text. The real-world trials are one robot and three routes, purely qualitative — a sanity check, not evidence.
One tension worth flagging: static expert replay improves measured WM fidelity on both distributions in Table IV yet lowers policy SR by 2.08 in Table III, so measured fidelity does not predict policy benefit. That weakens the stated mechanism (refined imagined feedback -> better policy), though it doesn't contradict the headline itself.
This is for anyone working on world-model-based policy optimization or embodied navigation post-training. It deserves serious refereeing despite the gaps; the right outcome is major revision requiring R_nav's definition, the baseline rewards, and sensitivity analysis — not a desk reject.
The strongest objection
The load-bearing assertion is that the recursive loop, not merely world-model feedback, drives the gains: §IV.B reports +6.57/5.60 (seen) and +7.94/6.50 (unseen) SR/SPL over GRPO-WM, with Table III tracing 80.31→88.25 through policy-induced replay and the self-curriculum. Every one of these policy updates is driven by r_i = R_nav(s, a_i, ô^+_i) in §III.B (Eqs. 2–3), yet R_nav is never defined: no formula, no inputs beyond its signature, no ablation, no validation. This is load-bearing because (1) all controlled comparisons and ablation rows are differences of training under an unspecified objective, so the measured deltas cannot be attributed to the claimed mechanisms; (2) the policy-only PPO/GRPO baselines are said to 'optimize the same navigation objective without action-conditioned visual feedback' (§IV.A), but R_nav as written consumes imagined frames from COSMOS — the paper never states what reward those baselines actually use, so either the reward differs across baselines (confounding GRPO vs GRPO-WM) or it is computable without WM outputs (making the imagined channel decorative and its causal role unverified); (3) the reader's failure mode is live: a WM adapted to the policy's proposals can make imagined rewards progressively more favorable without improving real outcomes, and nothing in the paper rules this out. The closed-loop metrics certify the full pipeline, not the imagined-feedback channel. A supporting tension: Table III row 2 (static expert replay) lowers SR by 2.08 relative to frozen, although Table IV shows static-expert adaptation improves measured WM fidelity on both evaluation sets (L1 0.1396→0.1352 expert; 0.1521→0.1505 policy-induced). Measured fidelity therefore does not predict policy benefit, and the asserted mechanism — refined imagined feedback → better policy — has no demonstrated correlate. The authors acknowledge the fixed-offline-dataset scope but do not acknowledge the undefined reward.
I reviewed the full text with the reader's verdict in hand. The arithmetic in Tables I–III checks out (GRPO-WM→RIWANAV +6.57/+5.60 seen, +7.94/+6.50 unseen; ablation 80.31→88.25; policy-induced replay +6.60; curriculum +3.42), the controlled-comparison protocol is well designed (same policy architecture, initialization, data, and evaluation protocol across frameworks), and the acknowledged limitations (fixed offline dataset, limited counterfactual exploration, qualitative real-world evidence) are stated honestly. Credit is also due for the Table IV/Fig. 3 fidelity measurements and the Table III ablation, whose rows do isolate the added components. The single most load-bearing gap remains what the reader identified: R_nav is the objective behind every policy update, every baseline, and every ablation delta, and it is never specified in §III.B. I sharpened this into a concrete causal threat: because the WM is itself adapted to the policy's proposals, an unspecified R_nav could be progressively satisfied by imagined frames rather than by genuine goal-ward behavior, and the closed-loop SR improvements cannot distinguish these explanations. I also note an internal tension that strengthens the concern: static-expert WM adaptation improves measured fidelity on both evaluation distributions (Table IV) while lowering navigation SR by 2.08 points (Table III row 2), so world-model fidelity per se is not the operative channel, and the paper does not address this. That said, I do not see evidence of a false result: the deltas are real reported metrics, the ablation logic supports the causal direction, and the failure modes I outline remain hypotheses pending the missing definition. CONDITIONAL therefore stands: the paper should be accepted only with R_nav specified and validated against ground-truth futures, ideally via the rank-correlation and oracle-credit checks above. No change to the reader's verdict is warranted.
The referee's report
The headline numbers are externally grounded: closed-loop navigation is measured against the real URBANNAV environment, and world-model fidelity against held-out state-action pairs. The controlled comparison in Table I fixes architecture, initialization, data, and protocol, and the arithmetic of every delta I checked (Table I deltas, Table III ablation rows 84.83−78.23=6.60 and 88.25−84.83=3.42, Table IV improvements) is consistent. If the central claim holds, the paper's contribution is a concrete instantiation of policy--world-model co-adaptation in visual navigation, with a transparent ablation decomposing the gain. The main risk is that the policy-improvement signal (the reward used by every policy update and every baseline) is never defined, which undermines attribution of the gains to the imagined-feedback mechanism rather than to some property of the unspecified scoring function.
Pith's reading of the paper (a machine's summary, not the authors')
RIWANAV claims that an action-conditioned world model and a navigation policy can be made to improve each other in a closed recursive loop, and that the recursion itself — not merely the presence of imagined feedback — is what produces the gains. The framework alternates two updates per cycle: the world model predicts visual futures for candidate actions, giving a GRPO policy update something to compare, and the improved policy then offers expert-consistent action–video pairs, prioritized by behavioral novelty and prediction error, to retrain the world model for the next round. On URBANNAV test-unseen the loop raises success rate from 80.31% (frozen world model) to 88.25%, and the ablation shows that naive retraining on expert data alone actually hurts (78.23%) — only retraining on the policy's own actions helps. If correct, this gives city-scale navigation a post-training recipe that learns from imagined failures at no physical trial cost, a middle path between imitation learning and real-environment reinforcement learning.
What carries the argument. RIWANAV's load-bearing object is the recursive self-improvement loop alternating two updates. In the policy phase, the COSMOS action-conditioned world model predicts future frames $\hat{o}_{t+1:t+T} = f_\phi(o_t, A(a), \ell)$ from a candidate waypoint chunk $a$ converted to a relative-pose sequence; the navigation scoring function $R_{\mathrm{nav}}$ turns those imagined futures into rewards for $M$ candidate actions sampled from the policy, and Group Relative Policy Optimization (GRPO) converts group-relative advantages into the policy update, restrained by ratio clipping, a KL penalty to the frozen initial policy $\pi_{\theta_0}$, and an expert-anchor loss. In the world-model phase, the improved policy's proposals are gated by the expert-consistency score $Q_i = \sum_{m\in\mathcal{M}} w_m \exp(-d_m(a_i, a^D_t)/\tau_m) / \sum_{m\in\mathcal{M}} w_m$ built from ADE, MADE, AOE, and MAOE discrepancies, and the grounded self-curriculum ranks admitted pairs by priority $p_i = Q_i(\epsilon_p + \lambda_N N_i + \lambda_E E_i)$, where novelty $N_i = 1 - \exp(-D_i/\tau_N)$ measures deviation from the initial policy's mean action and $E_i = 1 - \exp(-\tau_E e_i)$ measures current frame-wise L1 prediction error. Sampling without replacement in proportion to $p_i^\alpha$ feeds the world model's native video objective, and the loop repeats over $K = 19$ world-model update cycles, with the refined model supplying the imagined feedback for the next policy phase.
Core claim. The paper's central claim is that closing a recursive feedback loop between a policy and a world model yields navigation gains that frozen world-model feedback does not. Relative to GRPO-WM (the same optimizer with a frozen COSMOS world model), RIWANAV improves SR/SPL by 6.57%/5.60% on test-seen and 7.94%/6.50% on test-unseen. The recursive mechanism is built from two coupled updates: (1) the world model $f_\phi$ predicts visual outcomes $\hat{o}^+_i$ of $M$ sampled actions, a navigation score $R_{\mathrm{nav}}$ assigns each candidate a reward $r_i = R_{\mathrm{nav}}(s, a_i, \hat{o}^+_i)$, and a group-relative advantage $A_i = (r_i - \bar{r})/\mathrm{std}(r_{1:M}) + \epsilon_A$ drives a clipped policy update with a reference-policy KL penalty and an expert-anchor loss; (2) the improved policy proposes candidate actions, only those locally consistent with recorded expert motion (via the composite score $Q_i$ over ADE, MADE, AOE, and MAOE discrepancies) are paired with recorded futures, and a grounded self-curriculum prioritizes them by priority $p_i = Q_i(\epsilon_p + \lambda_N N_i + \lambda_E E_i)$ for retraining the world model on its native video objective. The ablation traces the gain through these components: static expert replay underperforms the frozen baseline (78.23% vs 80.31%), policy-induced replay with uniform sampling reaches 84.83%, and the grounded self-curriculum reaches 88.25%. The adapted world model also lowers L1/LPIPS on both expert and policy-induced action distributions, showing that the recursion improves feedback for the evolving policy without sacrificing prediction fidelity on the original expert-action distribution.
The premise it rests on. Every policy-improvement signal in the loop flows from the navigation scoring function $R_{\mathrm{nav}}$, which the paper never defines; if $R_{\mathrm{nav}}$ rewards predicted goal-ward trajectories without penalizing visual artifacts, the GRPO updates optimize a proxy that the co-adapted world model, itself trained on the policy's own proposals, can make progressively more favorable over successive rounds.
If the paper is right (Pith's inferences, not the authors')
Retraining a world model on expert action–video pairs alone is worse than leaving it frozen (78.23% vs 80.31% SR on test-unseen), so the value of adaptation comes specifically from the policy's own expert-consistent actions, not from more expert data.
Because the gains appear together on SR, SPL, NE, nDTW, SDTW, and OSR, policy-guided world-model adaptation improves the full rollout rather than only its stopping decision.
The adapted world model lowers L1/LPIPS on both expert actions and actions generated by the improved policy, with most of the improvement retained through the final recursive round, so the loop's feedback quality rises as the policy evolves instead of degrading.
The scheme runs entirely offline on collected video with no physical trial-and-error, and a policy trained this way guides a physical robot to language-specified landmarks, suggesting the post-training recipe transfers beyond closed-loop benchmarking.
The ablation ordering (static expert replay < frozen < policy replay < full curriculum) suggests a general principle the paper frames as recursive coupling: a world model must be built on the behavior distribution of the policy it will evaluate, which extends the on-policy/off-policy distinction from policy learning to model learning.
Because R_nav is applied only to imagined outcomes and never defined, the loop's stability depends on R_nav tracking true navigation success; a check the paper does not run is measuring the correlation between R_nav scores and realized success on held-out rollouts, or auditing which of the policy's failure modes R_nav actually catches.
The fixed expert-consistency threshold qmin = 0.35 bounds the loop's reach: if the improved policy's most useful actions drift beyond expert motion as training proceeds, the filter would reject the samples that most need world-model support, so an adaptive threshold that varies across recursive rounds is a concrete testable extension.
The single-domain evaluation makes generality provisional; the same recursive coupling could plausibly transfer to other video-conditioned policy problems with an action-conditioned future predictor, but the paper establishes nothing about that transfer.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Long-horizon urban navigation requires sequential local decisions whose errors can compound over time. Imitation learning (IL) rarely learns from failures, while physical trial-and-error reinforcement learning (RL) is costly. Action-conditioned world models can provide imagined feedback by predicting visual consequences for candidate actions. However, a frozen world model may become less reliable as the policy evolves. In this paper, we introduce RIWANAV, a post-training framework that casts the coupled adaptation of a world model and an action model (policy) as task-specific recursive self-improvement (RSI). Each cycle alternates two updates. The world model evaluates policy actions through imagined outcomes, providing comparative feedback for group-relative policy optimization (GRPO). The improved policy then constructs a grounded self-curriculum, selecting expert-consistent action-video pairs by behavioral novelty and prediction error. The refined world model supplies feedback for the next policy update, closing the recursive self-improvement loop. Experiments show that RIWANAV outperforms training baselines and prior methods, validating the proposed recursive self-improvement loop between the policy and world model. Real-world trials further demonstrate its practical applicability.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-07
The authors' main claim, as the classifier picked it out 1.00
“Experiments show that RIWANAV outperforms training baselines and prior methods, validating the proposed recursive self-improvement loop between the policy and world model.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.18 · human participants 0.10 · controlled comparison 0.90 · randomized assignment 0.02 · states quantitative result 0.04 · compares with prior results 0.97 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Long-horizon urban navigation requires sequential local decisions whose errors can compound over time. | background 1.00 | 0.05 | 0.02 | asserted 0.99 |
| Imitation learning (IL) rarely learns from failures, while physical trial-and-error reinforcement learning (RL) is costly. | background 0.57 | 0.07 | 0.02 | asserted 0.99 |
| Action-conditioned world models can provide imagined feedback by predicting visual consequences for candidate actions. | background 0.99 | 0.06 | 0.02 | hedged possibility 0.56 |
| However, a frozen world model may become less reliable as the policy evolves. | gap 0.61 | 0.08 | 0.02 | hedged possibility 0.72 |
| In this paper, we introduce RIWANAV, a post-training framework that casts the coupled adaptation of a world model and an action model (policy) as task-specific recursive self-improvement (RSI). | contribution 0.98 | 0.87 | 0.10 | asserted 1.00 |
| Each cycle alternates two updates. | method 0.99 | 0.65 | 0.02 | asserted 0.98 |
| The world model evaluates policy actions through imagined outcomes, providing comparative feedback for group-relative policy optimization (GRPO). | method 1.00 | 0.63 | 0.05 | asserted 0.99 |
| The improved policy then constructs a grounded self-curriculum, selecting expert-consistent action-video pairs by behavioral novelty and prediction error. | method 1.00 | 0.77 | 0.03 | asserted 0.99 |
| The refined world model supplies feedback for the next policy update, closing the recursive self-improvement loop. | method 1.00 | 0.64 | 0.11 | asserted 0.99 |
| Experiments show that RIWANAV outperforms training baselines and prior methods, validating the proposed recursive self-improvement loop between the policy and world model. | result 0.99 | 0.96 | 0.86 | asserted 1.00 |
| Real-world trials further demonstrate its practical applicability. | result 0.95 | 0.88 | 0.06 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-07: 1 request, 7,976 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
References (38)
- [1] X. Liu, J. Li, Y . Jiang, N. Sujay, Z. Yang, J. Zhang, J. Abanes, J. Zhang, and C. Feng, “CityWalker: Learning embodied urban navigation from web-scale videos,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 6875–6885 on Pith
- [2] Y . Mei, Y . Yang, L. Guo, Q. Wang, M.-M. Yu, X. He, W. Wu, and J. Liu, “UrbanNav: Learning language-guided embodied urban navigation from web-scale human trajectories,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 22, 2026, pp. 18 505–18 513 on Pith
- [3] F. Zhu, Z. Yan, Z. Hong, Q. Shou, X. Ma, and S. Guo, “WMPO: World model-based policy optimization for vision-language-action models,” inInternational Conference on Learning Representations, vol. 2026, 2026, pp. 62 486–62 502 on Pith
- [4] A. Bar, G. Zhou, D. Tran, T. Darrell, and Y . LeCun, “Navigation world models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 15 791–15 801 on Pith
- [5] S. Ye, Y . Ge, K. Zheng, S. Gao, S. Yuet al., “World action models are zero-shot policies,”arXiv preprint arXiv:2602.15922, 2026 on PithPith's reviewarXiv
- [6] T. Yuan, Z. Dong, Y . Liu, and H. Zhao, “Fast-WAM: Do world action models need test-time future imagination?”arXiv preprint arXiv:2603.16666, 2026 on PithPith's reviewarXiv
- [7] J. Schmidhuber, “Goedel machines: Fully self-referential optimal universal self-improvers,” inArtificial General Intelligence, B. Goertzel and C. Pennachin, Eds. Springer, 2007, pp. 199–226 on Pith
- [8] Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . K. Li, Y . Wu, and D. Guo, “DeepSeekMath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024 on PithPith's reviewarXiv
- [9] D. Shah, B. Eysenbach, N. Rhinehart, and S. Levine, “Rapid exploration for open-world navigation with latent goal models,” inProceedings of the 5th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, vol. 164, 2022, pp. 674–684 on Pith
- [10] D. Shah, A. Sridhar, N. Dashora, K. Stachowicz, K. Black, N. Hirose, and S. Levine, “ViNT: A foundation model for visual navigation,” in Proceedings of the 7th Conference on Robot Learning, ser. Proceedings of Machine Learning Research, vol. 229, 2023, pp. 711–733 on Pith
- [11] 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. [Online]. Available: https://arxiv.org/abs/2310.07896 on PithPith's reviewarXiv
- [12] N. Hirose, C. Glossop, D. Shah, and S. Levine, “OmniVLA: An omni-modal vision-language-action model for robot navigation,” in IEEE International Conference on Robotics and Automation (ICRA), on Pith
- [13] A. Zhang, X. Meng, L. Calliari, D.-K. Kim, S. Omidshafiei, J. Biswas, A. Agha, and A. Shaban, “VENTURA: Adapting image diffusion models for unified task conditioned navigation,”arXiv preprint arXiv:2510.01388, 2025 on PitharXiv
- [14] K. Chua, R. Calandra, R. McAllister, and S. Levine, “Deep reinforce- ment learning in a handful of trials using probabilistic dynamics models,” inAdvances in Neural Information Processing Systems, vol. 31, 2018, pp. 4754–4765 on Pith
- [15] M. Janner, J. Fu, M. Zhang, and S. Levine, “When to trust your model: Model-based policy optimization,” inAdvances in Neural Information Processing Systems, vol. 32, 2019, pp. 12 498–12 509 on Pith
- [16] D. Hafner, T. Lillicrap, I. Fischer, R. Villegas, D. Ha, H. Lee, and J. Davidson, “Learning latent dynamics for planning from pixels,” in Proceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 97, 2019, pp. 2555–2565 on Pith
- [17] D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi, “Dream to control: Learning behaviors by latent imagination,” inInternational Conference on Learning Representations, 2020 on Pith
- [18] D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap, “Mastering diverse control tasks through world models,”Nature, vol. 640, pp. 647–653, 2025 on Pith
- [19] N. Hansen, H. Su, and X. Wang, “TD-MPC2: Scalable, robust world models for continuous control,” inInternational Conference on Learning Representations, vol. 2024, 2024, pp. 47 376–47 405 on Pith
- [20] V . Micheli, E. Alonso, and F. Fleuret, “Transformers are sample- efficient world models,” inInternational Conference on Learning Representations, 2023 on Pith
show 18 more
- [21] N. Hansen, Y . Lin, H. Su, X. Wang, V . Kumar, and A. Rajeswaran, “MoDem: Accelerating visual model-based reinforcement learning with demonstrations,” inInternational Conference on Learning Representa- tions, 2023 on Pith
- [22] G. Zhou, H. Pan, Y . LeCun, and L. Pinto, “DINO-WM: World models on pre-trained visual features enable zero-shot planning,” in Proceedings of the 42nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 267, 2025, pp. 79 115–79 135 on Pith
- [23] X. Yao, J. Gao, and C. Xu, “NavMorph: A self-evolving world model for vision-and-language navigation in continuous environments,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 5536–5546 on Pith
- [24] J. Yang, K. Lin, J. Li, W. Zhang, T. Lin, L. Wu, Z. Su, H. Zhao, Y .-Q. Zhang, L. Chen, P. Luo, X. Yue, and H. Li, “RISE: Self-improving robot policy with compositional world model,”Proceedings of Robotics: Science and Systems, 2026 on Pith
- [25] R. Wang, J. Lehman, J. Clune, and K. O. Stanley, “POET: Open- ended coevolution of environments and their optimized solutions,” in Proceedings of the Genetic and Evolutionary Computation Conference, 2019, pp. 142–151 on Pith
- [26] T. Yu, G. Thomas, L. Yu, S. Ermon, J. Zou, S. Levine, C. Finn, and T. Ma, “MOPO: Model-based offline policy optimization,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 14 129– 14 142 on Pith
- [27] R. Kidambi, A. Rajeswaran, P. Netrapalli, and T. Joachims, “MOReL: Model-based offline reinforcement learning,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 21 810–21 823 on Pith
- [28] X. Wang, W. Wongkamjan, R. Jia, and F. Huang, “Live in the moment: Learning dynamics model adapted to evolving policy,” in Proceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 202, 2023, pp. 36 470–36 493 on Pith
- [29] K. Kim, M. Sano, J. D. Freitas, N. Haber, and D. Yamins, “Active world model learning with progress curiosity,” inProceedings of the 37th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 119, 2020, pp. 5306–5315 on Pith
- [30] R. Sekar, O. Rybkin, K. Daniilidis, P. Abbeel, D. Hafner, and D. Pathak, “Planning to explore via self-supervised world models,” inProceedings of the 37th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 119, 2020, pp. 8583– 8592 on Pith
- [31] I. Kauvar, C. Doyle, L. Zhou, and N. Haber, “Curious replay for model- based adaptation,” inProceedings of the 40th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 202, 2023, pp. 16 018–16 048 on Pith
- [32] NVIDIAet al., “Cosmos 3: Omnimodal world models for physical AI,”arXiv preprint arXiv:2606.02800, 2026 on PithPith's reviewarXiv
- [33] M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov et al., “DINOv2: Learning robust visual features without supervision,” Transactions on Machine Learning Research, 2024 on Pith
- [34] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” inProceedings of the 38th International Conference on Machine Le... on Pith
- [35] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,”arXiv preprint arXiv:1707.06347, 2017 on PithPith's reviewarXiv
- [36] 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,”arXiv preprint arXiv:1807.06757, 2018 on PithPith's reviewarXiv
- [37] G. Ilharco, V . Jain, A. Ku, E. Ie, and J. Baldridge, “General evaluation for instruction conditioned navigation using dynamic time warping,”CoRR, vol. abs/1907.05446, 2019. [Online]. Available: https://arxiv.org/abs/1907.05446 8 on PithPith's reviewarXiv
- [2026] [Online]. Available: https://arxiv.org/abs/2509.19480 on PitharXiv
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment