Beckmann 世界模型：面向一步动作条件预测的终端映射学习
原标题：Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction
AI 摘要
论文提出 Beckmann World Models（BWM），将 Beckmann Transport Models 的终端映射守恒律适配到动作条件视频预测，无需预训练生成教师即可从噪声、视觉历史和动作直接一步预测未来帧。作者指出传输一致性无法保证纯噪声输入下的预测精度，因此引入端点对齐训练，直接监督源预测并强调运动与空间结构。在 PushT 和 Robomimic Can 上，最终模型相比同骨干回归基线分别降低 64 帧预测 MSE，并在策略评估中取得最低平均 IoU 误差。
正文节选
Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction Abstract Action-conditioned world models support planning and policy evaluation by predicting future states. However, learning an accurate one-step predictor can require substantial training computation. We introduce Beckmann World Models (BWM), which learn a direct map from noise, visual history, and actions to future frames without a pretrained generative teacher. We adapt the terminal-map conservati