Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning

arXiv 26.3 2026 69.2 method

TLDR

RL post-training stabilizes autoregressive robot world model rollouts, achieving SOTA on DROID dataset.

Reasoning

The paper introduces a novel RL-based fine-tuning method for diffusion world models, addressing error accumulation in multi-step video prediction. Strengths include a principled contrastive RL objective and strong empirical gains on real robot data. Weaknesses are the narrow domain focus and lack of discussion on generalization to other tasks or environments.

Read-first score

Read-first score 69.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 8%

100

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

Topical relevance 42%

81.4

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 152.

Keyword Scores

world model

generative world model

video world model

interactive world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

Deep Analysis

Innovations

Reinforcement learning post-training scheme that trains the world model on its own autoregressive rollouts rather than ground-truth histories
Training protocol that generates and compares multiple candidate variable-length futures from the same rollout state, reinforcing higher-fidelity predictions
Efficient, multi-view visual fidelity rewards combining complementary perceptual metrics across camera views, aggregated at clip level for dense, low-variance training signal
New state-of-the-art for rollout fidelity on the DROID dataset, outperforming strongest baseline on all metrics

Methodology

The approach adapts a recent contrastive RL objective for diffusion models to a world model setting, applying it as a post-training scheme where the model is trained on its own autoregressive rollouts. A training protocol generates multiple candidate variable-length futures from the same rollout state and uses multi-view visual fidelity rewards—combining complementary perceptual metrics across camera views and aggregated at the clip level—to reinforce higher-fidelity predictions over lower-fidelity ones.

Key Results

On the DROID dataset, the method reduces LPIPS by 14% on external cameras, improves SSIM by 9.1% on the wrist camera, wins 98% of paired comparisons, and achieves an 80% preference rate in a blind human study.

Resources

Paper