Problem

While VLA models perform strongly in language-conditioned robotic manipulation, their robustness to language variants such as multilingual instructions is poorly understood, and existing unified alignment methods have limited effectiveness.

Method Overview

• Backbone: Vision-Language-Action (VLA) 模型
• Key module: 步骤级语言敏感性评分模块、推理时选择性干预机制
• Loss: 策略损失（Policy Loss），干预阶段采用表示对齐策略

Takeaways

Personal understanding — corrections welcome

Experiments indicate that the language robustness of VLA models is essentially a step-wise control problem, where language bias concentrates on specific critical steps rather than being uniformly distributed.

Limitations & Questions

The evaluation is primarily based on the LIBERO simulation benchmark, requiring verification of generalization in real robotic scenarios; the intervention method relies on offline English data to construct sensitivity standards.