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

AI 导读

AD-WM通过动作判别性世界模型提升反事实模型预测控制效果，在多个仿真环境和真实任务中显著提高成功率。

AD-WM introduces action-discriminative world models for counterfactual MPC, achieving significant success rate improvements in simulations and zero-shot transfer tasks.

重点速览

• 传统世界模型无法区分候选动作，AD-WM通过残差动态与动作恢复正则化解决此问题 AD-WM addresses action discrimination gap via residual dynamics and action recovery regularization
• 在OGBench-Cube测试中，成功率从3.7%提升至52.0%，超越基线模型 Achieves 52.0% success rate on OGBench-Cube vs 3.7% baseline
• 零样本迁移至Franka机械臂时，抓取成功率从42.2%提升至71.1% Improves zero-shot transfer success to Franka from 42.2% to 71.1%
• 理论基于条件互信息的归一化恢复目标，保留动作相关差异信息 Theoretical foundation based on conditional mutual information normalization
• 规划诊断显示CEM对齐的精英遗憾更准确反映成功概率 CEM-aligned elite regret better correlates with success probability

一句话：世界模型需保留动作差异信息以支持反事实决策，而非单纯优化事实预测精度。 / World models for planning must preserve action-dependent differences for counterfactual selection.

推荐理由 该方法为强化学习中的反事实决策提供了新范式，具有实际部署价值。