预测动力学能否存在于物理世界中？

摘要

预测性物理人工智能系统输出状态展开、动作片段和潜在规划，但低均方根误差（RMSE）并不意味特定方案在物理上可执行。我们将物理可容许性定义为预测-控制接口：在执行前，将解码后的方案视为候选动力学特性，并通过运动学、动力学以及直接到复合的视界条件进行评估。通过并不代表任务成功；拒绝则表明违反了特定物理包络，并提供组件层面的理由。在Hugging Face LeRobot PushT上，受控伪造实验表明，单步预测RMSE与标准化动力学残差达到受试者工作特征曲线下面积（AUC）0.982和0.972，仅运动学条件达到AUC 0.592，而完整门控达到AUC 0.957，并附带条件级归因。在基于回放的中介实验中，残差滤波器与完整物理可容许性门控可阻止87%-89%的无效方案，同时保持平均进度接近0.998。

English

Predictive Physical AI systems output state rollouts, action chunks, and latent plans, yet a low root-mean-square error (RMSE) does not imply that a particular proposal is physically executable. We formulate physical admissibility as a prediction-control interface: before execution, a decoded proposal is treated as candidate dynamics and evaluated using kinematic, dynamic, and direct-to-composed horizon conditions. Passing is not a certificate of task success; rejection identifies violation of the specified physical envelope and gives a component-level reason. On Hugging Face LeRobot PushT, controlled falsification shows that one-step prediction-RMSE and standardized dynamics residuals reach area under the receiver operating characteristic curve (AUC) 0.982 and 0.972, kinematic-only conditions reach AUC 0.592, and the full gate reaches AUC 0.957 with condition-level attribution. In replay-based intervention experiments, residual-based filters and the full physical-admissibility gate prevent 87-$89% of invalid proposals while preserving mean progress near 0.998.