CLaD：通过跨模态潜变量动力学进行有据前瞻的规划
CLaD: Planning with Grounded Foresight via Cross-Modal Latent Dynamics
把本体感知与语义状态用非对称交叉注意力对齐成统一的潜空间动力学，并用自监督预测未来潜变量来条件化扩散策略；做视觉-动作联合建模或机器人操作策略学习时，这套『预测有据的潜变量前瞻 + 扩散动作生成』结构可直接复用。
原文依据 asymmetric cross-attention that allows kinematic transitions to query semantic ones
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论机器人操作涉及运动学与语义层面的状态转移，而这些转移通过底层动作内在地耦合在一起。然而，现有方法要么在语义空间中规划，要么在潜空间中规划，并未显式对齐这些跨模态的状态转移。为解决这一问题，我们提出 CLaD，一个通过非对称交叉注意力建模本体感知状态与语义状态如何在动作作用下共同演化的框架，该注意力允许运动学转移查询语义转移。CLaD 借助带 EMA 目标编码器的自监督目标与辅助重建损失来预测有据的潜变量前瞻，从而在防止表征坍缩的同时将预测锚定到可观测状态。预测得到的前瞻与观测相调制，用以条件化扩散策略以生成动作。在 LIBERO-LONG 基准上，CLaD 取得 94.7% 的成功率，以显著更少的参数量与大型 VLA 模型相比具有竞争力。
英文原文摘要 919 字符
Robotic manipulation involves kinematic and semantic transitions that are inherently coupled via underlying actions. However, existing approaches plan within either semantic or latent space without explicitly aligning these cross-modal transitions. To address this, we propose CLaD, a framework that models how proprioceptive and semantic states jointly evolve under actions through asymmetric cross-attention that allows kinematic transitions to query semantic ones. CLaD predicts grounded latent foresights via self-supervised objectives with EMA target encoders and auxiliary reconstruction losses, preventing representation collapse while anchoring predictions to observable states. Predicted foresights are modulated with observations to condition a diffusion policy for action generation. On LIBERO-LONG benchmark, CLaD achieves 94.7% success rate, competitive with large VLAs with significantly fewer parameters.