Simulation-to-real validation of paraphrase-robustness vulnerabilities

Determine whether the paraphrase-robustness vulnerabilities observed for vision–language–action models when evaluated with LIBERO-Para in the LIBERO simulation environment persist when these models are deployed on physical robotic platforms.

Background

All experiments and analyses in the paper are conducted in the LIBERO simulation environment using the proposed LIBERO-Para benchmark and the PRIDE metric. The authors find substantial performance degradation under paraphrased instructions and attribute failures largely to planning-level errors driven by object-referent variation.

However, simulations may not fully reflect real-world factors such as sensor noise, physics, and rendering fidelity. The authors explicitly note that additional validation is necessary to ascertain whether the same paraphrase-robustness vulnerabilities documented in simulation also occur on physical robots.