AttenA+: Rectifying Action Inequality in Robotic Foundation Models

现有机器人基础模型虽然强大，但隐含地假设时间上是同质的：在优化时把所有动作都视为同等有信息量。这种继承自语言建模的“平坦”训练范式，对操作任务底层的物理层级并不敏感。现实中，机器人轨迹本质上是异质的：低速片段往往通过高精度交互决定任务成功，而高速运动通常只是容错性较强的过渡。统一 loss 权重与物理关键性之间的错配，从根本上限制了当前 Vision-Language-Action（VLA）模型和 World-Action Model（WAM）在复杂长程任务中的表现。为纠正这一问题，作者提出 AttenA+，一个架构无关的框架，通过速度驱动的 action attention 优先学习运动学上关键的片段。AttenA+ 基于逆速度场重加权训练目标，使模型学习容量自然对齐操作任务的物理需求。作为即插即用增强，它可以无结构修改、无额外参数地集成到现有骨干中。大量实验表明，AttenA+ 显著提升当前 SOTA 模型的上限：在 Libero benchmark 上将 OpenVLA-OFT 提升到 98.6%（+1.5%），在 RoboTwin 2.0 上将 FastWAM 提升到 92.4%（+0.6%）。真实 Franka 机械臂验证进一步展示了其鲁棒性和跨任务泛化能力。该工作说明，挖掘动作序列的内在结构先验，是标准 scaling law 之外一种高效、物理感知的补充路径。

Existing robotic foundation models, while powerful, are predicated on an implicit assumption of temporal homogeneity: treating all actions as equally informative during optimization. This "flat" training paradigm, inherited from language modeling, remains indifferent to the underlying physical hierarchy of manipulation. In reality, robot trajectories are fundamentally heterogeneous, where low-velocity segments often dictate task success through precision-demanding interactions, while high-velocity motions serve as error-tolerant transitions. Such a misalignment between uniform loss weighting and physical criticality fundamentally limits the performance of current Vision-Language-Action (VLA) models and World-Action Models (WAM) in complex, long-horizon tasks. To rectify this, we introduce AttenA+, an architecture-agnostic framework that prioritizes kinematically critical segments via velocity-driven action attention. By reweighting the training objective based on the inverse velocity field, AttenA+ naturally aligns the model's learning capacity with the physical demands of manipulation. As a plug-and-play enhancement, AttenA+ can be integrated into existing backbones without structural modifications or additional parameters. Extensive experiments demonstrate that AttenA+ significantly elevates the ceilings of current state-of-the-art models. Specifically, it improves OpenVLA-OFT to 98.6% (+1.5%) on the Libero benchmark and pushes FastWAM to 92.4% (+0.6%) on RoboTwin 2.0. Real-world validation on a Franka manipulator further showcases its robustness and cross-task generalization. Our work suggests that mining the intrinsic structural priors of action sequences offers a highly efficient, physics-aware complement to standard scaling laws, paving a new path for general-purpose robotic control.