Skeleton-Grounded Latent Action Model for Robot World Models

arXiv 2026 38.3 method, application

TLDR

SkeLAM decomposes latent actions into skeleton-guided kinematic and residual components, enabling controllable robot world models that generate future observations from actions.

评分理由

The paper introduces a novel skeleton-guided latent action decomposition with cross-modal swap reconstruction and adapter-based action conditioning, supported by strong DROID benchmark results and zero-shot transfer. Weaknesses include limited discussion of limitations and the abstract truncating policy evaluation details.

Read-first 评分解释

综合优先阅读分 38.3，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 53。

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%

75.7

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

方法质量 15%

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：评估、结果

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：无

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

排序敏感性

稳定性：volatile；排名波动范围：498。

关键词评分

world model

world dynamics prediction

generative world model

interactive world model

video world model

world simulator

model-based reinforcement learning world model

深度分析

创新点

Decomposes latent actions into a skeleton-grounded kinematic component and a residual component to separate robot motion from complementary visual dynamics.
Uses cross-modal swap reconstruction to encourage RGB-derived and skeleton-derived kinematic representations to be interchangeable during video reconstruction.
Introduces separate conditioning pathways and branch dropout to reduce dependence on future-derived residual information.
Maps robot actions to kinematic space via a lightweight adapter, followed by world-model fine-tuning for controllable generation.

方法

SkeLAM learns robot world models from videos on DROID by decomposing latent actions into skeleton-supervised kinematic and residual components. It uses cross-modal swap reconstruction, separate conditioning pathways, and branch dropout, then maps robot actions to kinematic space with a lightweight adapter and fine-tunes the world model. Evaluation compares against DreamDojo using end-effector ADE and LPIPS, with motion probing, qualitative transfer, and real-robot policy success agreement.

关键结果

On DROID, SkeLAM reduces end-effector ADE by 19.5% and LPIPS by 11.1% compared with DreamDojo. Motion probing supports kinematic specialization, qualitative results demonstrate zero-shot human-to-robot motion transfer, and policy evaluation shows agreement between predicted and real-robot success rates.

资源

论文