DeltaWorld: Physically Consistent Interactive World Simulators via Action-Conditioned Latent Increment Learning

arXiv 2026 46.2 method

TLDR

DeltaWorld introduces action-conditioned latent increment learning and interaction-aware alignment for physically consistent interactive world simulators, reducing FVD and LPIPS on manipulation benchmarks.

评分理由

The paper addresses a clear limitation of existing world models by predicting action-induced latent changes, and it evaluates on a benchmark plus a cross-robot dataset. However, the abstract is truncated and lacks details on baselines, ablations, and limitations, making full assessment difficult.

Read-first 评分解释

综合优先阅读分 46.2，由主题、引用、图谱、方法、可复现性和近期性等信号加权得到。原始总分保留为 55。

方法质量 15%

100

检查可见的摘要与分析字段，寻找实验、数据集、基线、指标和局限性等方法证据。命中信号：基线、基准、数据集、评估、指标、结果

近期性 5%

100

使用温和的时间衰减，让近期论文更容易浮现，同时保留较早基础工作的价值。年份：2026

主题相关性 25%

78.6

使用现有 LLM 关键词相关性评分，并归一化到 0-100。关键词：world model、world simulator、generative world model、interactive world model、video world model、world dynamics prediction、model-based reinforcement learning world model

可复现性 15%

检查链接和可见文本中的论文、代码、数据集、工件与仓库信号。论文：有；代码：无；数据：无；命中信号：数据集

图谱声望 15%

5.7

基于可用 OpenAlex 引用关系构建局部引用图，并计算 PageRank 风格的图谱声望。局部 PageRank：0.0016

引用影响力 15%

使用 OpenAlex 形态的引用元数据作为文献关注度信号，并与论文本身质量分开处理。引用数：0

引用速度 10%

引用速度按发表年限估算年均引用，降低旧论文天然占优的偏差。年均引用：0.00

研究版图角色

前沿论文方法锚点

排序敏感性

稳定性：volatile；排名波动范围：798。

关键词评分

world simulator

interactive world model

world model

generative world model

world dynamics prediction

video world model

model-based reinforcement learning world model

深度分析

创新点

Delta Latent Transition Model (Delta-LTM) predicts action-induced latent feature changes and adds them to the current latent state instead of directly predicting the next latent state.
Interaction-aware Latent Alignment constructs counterfactual interaction regions and supervises interaction-related latent changes to reduce object interpenetration and excessive deformation.
DeltaWorld is presented as a physically consistent interactive world simulator for robotic manipulation across multiple robot embodiments and tasks.

方法

DeltaWorld uses action-conditioned latent increment learning, where the Delta-LTM predicts action-induced latent feature changes and adds them to the current latent state to produce the next state. Interaction-aware Latent Alignment constructs counterfactual interaction regions to supervise interaction-related latent changes. The method is evaluated on the IWS manipulation benchmark and a self-collected cross-robot dataset, using FVD and LPIPS metrics relative to the IWS baseline.

关键结果

On the cross-robot dataset, DeltaWorld reduces FVD by 46.6% and LPIPS by 31.1% relative to the IWS baseline, indicating improved long-horizon action-conditioned video prediction.

资源

论文