Learning Invariant Visual Representations for Planning with Joint-Embedding Predictive World Models

arXiv 26.2 2026 60.5 method

TLDR

Improves latent predictive world model robustness by adding bisimulation encoder to ignore slow features, tested on navigation with distractors.

Reasoning

The paper identifies a real limitation of JEPA models and proposes a principled solution using bisimulation. However, evaluation is limited to a simple navigation task with synthetic distractors, lacking real-world or complex benchmarks.

Read-first score

Read-first score 60.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark,evaluation

Topical relevance 42%

54.3

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 337.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

video world model

generative world model

world simulator

Deep Analysis

Innovations

Augmenting the predictive objective of joint-embedding predictive world models with a bisimulation encoder to enforce control-relevant state equivalence
Reducing latent space size up to 10x compared to DINO-WM while maintaining robustness to slow features
Demonstrating agnosticism to the choice of pretrained visual encoder (DINOv2, SimDINOv2, iBOT)

Methodology

The model extends a joint-embedding predictive architecture (JEPA) by incorporating a bisimulation encoder that maps states with similar transition dynamics to nearby latent representations, thereby suppressing slow features like background changes and distractors. It is evaluated on a simple navigation task under test-time background variations and visual distractors, with comparisons to the DINO-WM baseline.

Key Results

Across all benchmarks, the proposed model consistently improves robustness to slow features while operating in a latent space up to 10x smaller than that of DINO-WM.

Limitations

Evaluation is limited to a simple navigation task, leaving generalization to more complex environments unverified
The abstract does not discuss potential failure cases or sensitivity to hyperparameters of the bisimulation encoder
Robustness is only tested under specific types of background changes and distractors; broader or more extreme variations are not explored

Resources

Paper