Skip to content
Escape
Meridian

The future is still open.

Park, Mingyu, Noh, Samyeul, Myung, Hyun, Lee, Donghwan · Oct 2, 2026

Mingyu Park and coauthors on VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning

Markdown · JSON · Original source

The VIGOR paper’s authors argue that visual changes can undermine model-based (MBRL) twice: they shift encoder outputs, then errors compound during recursive latent rollouts. They propose VIGOR, a method intended to help systems trained in one visual setting handle unseen distractions such as background, lighting, or camera changes without additional training on those distractions. VIGOR combines asymmetric weak-to-strong augmentation, dynamics-level consistency, and encoder-level stabilization. The authors report tests using TD-MPC2 on the Control Suite (DMC) and Robosuite, where VIGOR outperformed the second-best baseline by 3.4% and 43.6%, respectively, in generalization. They also report sample efficiency comparable to TD-MPC2 and fewer environment interactions than other baselines. Ablations, the authors say, found strong generalization across distinct perturbation families, and a DreamerV3 implementation also improved results on both . The paper’s argument is that consistency in latent representations, rather than a particular augmentation choice, is central to its reported robustness. These results concern the reported benchmarks; the abstract does not establish how the method performs in other settings.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Oct 2, 2026 (2026-10-02).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-08T20:10:50.612Z. First observed: 2026-10-05T08:55:12.008Z. Retrieved: 2026-11-08T20:09:46.896Z.

Referenced by

Recorded citations and public file attachments. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The evidence supplied is the abstract, which reports results but does not provide the experimental details needed to assess the comparisons. The reported results concern DMC and Robosuite; the abstract does not report performance on other settings.

Open question

How does VIGOR perform on visual distractions outside the Control Suite and Robosuite ?

Source-stated

The authors say MBRL is vulnerable to visual distractions because they shift encoder outputs and the resulting errors compound through recursive latent rollouts.

distractions push encoder outputs out of distribution, and these errors compound through recursive latent rollouts.

Source-stated

The authors propose VIGOR to support zero-shot generalization to unseen visual distractions while retaining the backbone's sample efficiency.

enables zero-shot generalization to unseen distractions while retaining its backbone's sample efficiency.

Source-stated

VIGOR combines asymmetric weak-to-strong augmentation, dynamics-level consistency, and encoder-level stabilization.

three interdependent components: (i) asymmetric weak-to-strong augmentation, which pairs weak-only and weak-to-strong latent views within a batch; (ii) dynamics-level consistency, which enforces augmentation-invariant transition predictions via direct latent regression; and (iii) encoder-level

Source-stated

On DMC and Robosuite, the authors report that VIGOR instantiated on TD-MPC2 outperformed the second-best baseline by 3.4% and 43.6%, respectively, in zero-shot generalization.

surpassing the second-best baseline by 3.4% on DMC and 43.6% on Robosuite

Source-stated

The authors report that augmentations from distinct perturbation families preserve strong generalization, and that a DreamerV3 instantiation improves zero-shot generalization on both benchmarks.

augmentations from distinct perturbation families preserve strong generalization

Source-stated

The authors interpret their ablations as indicating that latent-space consistency, rather than the augmentation choice, drives robustness.

latent-space consistency, not the augmentation choice, drives robustness.

Source and contributor context

Park, Mingyu, Noh, Samyeul, Myung, Hyun, Lee, Donghwan · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Zero-shot

Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).

Reinforcement learning

Training by trial and error, in which a system is rewarded for the outcomes its designers want.

Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Google DeepMind

Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.

Editorial provenance

Version 1. Publication ID: 97df3d25-dbc7-45ca-819b-9202a5115cf3.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T20:10:50.612Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 05b5b83ff1a2faceea1132348e792e79d2a46e8becef6c986ac75bf8ae7a18bc. Source text hash: 70c98d5fb5162c2cd217f3e032b0daa884d63fff56a75bf8c7a060597a8dbdf5.

Revision and source-check history

Version 1 · 2026-11-08T20:10:50.612Z

Mingyu Park and coauthors on VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning

The VIGOR paper’s authors argue that visual changes can undermine model-based (MBRL) twice: they shift encoder outputs, then errors compound during recursive latent rollouts. They propose VIGOR, a method intended to help systems trained in one visual setting handle unseen distractions such as background, lighting, or camera changes without additional training on those distractions. VIGOR combines asymmetric weak-to-strong augmentation, dynamics-level consistency, and encoder-level stabilization. The authors report tests using TD-MPC2 on the Control Suite (DMC) and Robosuite, where VIGOR outperformed the second-best baseline by 3.4% and 43.6%, respectively, in generalization. They also report sample efficiency comparable to TD-MPC2 and fewer environment interactions than other baselines. Ablations, the authors say, found strong generalization across distinct perturbation families, and a DreamerV3 implementation also improved results on both . The paper’s argument is that consistency in latent representations, rather than a particular augmentation choice, is central to its reported robustness. These results concern the reported benchmarks; the abstract does not establish how the method performs in other settings.

Original source

cs.AI · Article

Source details

Original source date: Oct 2, 2026 (2026-10-02).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-08T20:10:50.612Z. First observed: 2026-10-05T08:55:12.008Z. Retrieved: 2026-11-08T20:09:46.896Z.

Limits

The evidence supplied is the abstract, which reports results but does not provide the experimental details needed to assess the comparisons. The reported results concern DMC and Robosuite; the abstract does not report performance on other settings.

Open question

How does VIGOR perform on visual distractions outside the Control Suite and Robosuite ?

Source-stated

The authors say MBRL is vulnerable to visual distractions because they shift encoder outputs and the resulting errors compound through recursive latent rollouts.

distractions push encoder outputs out of distribution, and these errors compound through recursive latent rollouts.

Source-stated

The authors propose VIGOR to support zero-shot generalization to unseen visual distractions while retaining the backbone's sample efficiency.

enables zero-shot generalization to unseen distractions while retaining its backbone's sample efficiency.

Source-stated

VIGOR combines asymmetric weak-to-strong augmentation, dynamics-level consistency, and encoder-level stabilization.

three interdependent components: (i) asymmetric weak-to-strong augmentation, which pairs weak-only and weak-to-strong latent views within a batch; (ii) dynamics-level consistency, which enforces augmentation-invariant transition predictions via direct latent regression; and (iii) encoder-level

Source-stated

On DMC and Robosuite, the authors report that VIGOR instantiated on TD-MPC2 outperformed the second-best baseline by 3.4% and 43.6%, respectively, in zero-shot generalization.

surpassing the second-best baseline by 3.4% on DMC and 43.6% on Robosuite

Source-stated

The authors report that augmentations from distinct perturbation families preserve strong generalization, and that a DreamerV3 instantiation improves zero-shot generalization on both benchmarks.

augmentations from distinct perturbation families preserve strong generalization

Source-stated

The authors interpret their ablations as indicating that latent-space consistency, rather than the augmentation choice, drives robustness.

latent-space consistency, not the augmentation choice, drives robustness.

Source and contributor context

Park, Mingyu, Noh, Samyeul, Myung, Hyun, Lee, Donghwan · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Zero-shot

Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).

Reinforcement learning

Training by trial and error, in which a system is rewarded for the outcomes its designers want.

Benchmark

A defined set of tests used to compare systems. Results depend on the tasks and measures selected.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Google DeepMind

Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.

Editorial provenance

Version 1. Publication ID: 97df3d25-dbc7-45ca-819b-9202a5115cf3.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-08T20:10:50.612Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 05b5b83ff1a2faceea1132348e792e79d2a46e8becef6c986ac75bf8ae7a18bc. Source text hash: 70c98d5fb5162c2cd217f3e032b0daa884d63fff56a75bf8c7a060597a8dbdf5.

passed · 2026-11-08T20:10:50.612Z · anthropic/claude-opus-5.5

Publication: 97df3d25-dbc7-45ca-819b-9202a5115cf3

attribution — pass: The reading attributes the paper to Mingyu Park and coauthors, matching the listed authors. It consistently frames the vulnerability argument, method and benchmark results as the authors' claims rather than established facts.

Authors:Mingyu Park, Samyeul Noh, Hyun Myung, Donghwan Lee

summary — pass: The summary faithfully conveys the two-level vulnerability argument, the three components, the TD-MPC2 results on DMC and Robosuite, the ablation interpretation and the DreamerV3 extension. It explains zero-shot as handling unseen distractions without additional training on them. Terms such as latent rollouts and augmentation are kept exactly, suitable for tooltips. The closing scope note and the question about other settings follow directly from the abstract.

MBRL suffers from a two-level vulnerability: distractions push encoder outputs out of distribution, and these errors compound through recursive latent rollouts.

date — pass: The originalPublished value is 2026-10-02, which matches the v1 submission date. It does not use the 5 Nov 2026 v2 revision or the November feed date.

Submitted on 2 Oct 2026 (v1), last revised 5 Nov 2026 (this version, v2)

support — pass: Each claim matches the abstract: the 3.4% and 43.6% margins over the second-best baseline, sample efficiency comparable to the backbone, fewer environment interactions than other baselines, and the augmentation-agnostic ablations. The DreamerV3 improvement on both benchmarks is also stated. The draft adds no unsupported benefits or forecasts, and its limitations correctly note that only the abstract is supplied.

surpassing the second-best baseline by 3.4% on DMC and 43.6% on Robosuite, with sample efficiency comparable to its backbone and fewer environment interactions than the other baselines

media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.