AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control

arXiv 2026 37.4 method

TLDR

Introduces AD-WM, an action-discriminative joint-embedding world model for counterfactual MPC, improving planning success in simulation and real robot transfer by preserving action information.

评分理由

The paper presents a strong methodological contribution with action-discriminative regularization and demonstrates clear empirical gains across simulation and real-world Franka transfer. Weaknesses include limited visible evidence on generative or video capabilities, and the auxiliary heads are discarded, so broader generality is not established from the abstract alone.

Read-first 评分解释

综合优先阅读分 37.4，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 38。

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

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：基线、结果

主题相关性 25%

54.3

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：代码、GitHub、复现

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：517。

关键词评分

world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

generative world model

video world model

深度分析

创新点

Introduces AD-WM, an action-discriminative joint-embedding world model for counterfactual model predictive control.
Combines residual latent dynamics with predictor-level action-recovery regularization.
Uses inverse dynamics and a normalized recovery objective motivated by conditional mutual information to preserve action information in planning transitions.
Auxiliary action-recovery heads are discarded at test time, leaving MPC unchanged.

方法

AD-WM is a latent world model with residual latent dynamics and auxiliary action-recovery heads trained using inverse dynamics and a normalized recovery objective. The auxiliary heads are removed at test time. It is evaluated on OGBench-Cube, five simulation environments against a matched LeWM/reproduced baseline, and zero-shot transfer to a Franka setup with a frozen V-JEPA 2 encoder and matched DROID post-training.

关键结果

AD-WM improves OGBench-Cube hard-start success from 3.7% to 52.0% over a matched LeWM baseline and improves mean success over the reproduced baseline in four of five simulation environments. CEM-aligned elite regret tracks closed-loop success more closely than factual prediction error or whole-bank action ranking, and zero-shot Franka basic pick-and-place success increases from 42.2% to 71.1%.

局限性

The reported improvement in simulation is not universal: mean success improved in four of five simulation environments.
The zero-shot transfer result is shown only for basic pick-and-place on the Franka setup.
The abstract does not state other explicit limitations or failure modes.

资源

论文