Dimensions overall score 7.0

• VLA models exhibit consistent performance degradation of 22-52 percentage points when tested with paraphrased instructions compared to original instructions.result 90%
• The primary driver of performance degradation under paraphrasing is object-level lexical variation, where even simple synonym substitutions cause large performance drops.result 90%
• 80-96% of task failures under paraphrased instructions arise from planning-level trajectory divergence, not from execution errors.result 90%
• The binary success rate metric is insufficient for evaluating paraphrase robustness because it treats all paraphrases equally and obscures whether models rely on easier cases.limitation 85%
• The paper introduces LIBERO-Para, a diagnostic benchmark for paraphrase robustness that independently varies action expressions and object references for fine-grained analysis.method 95%
• The paper proposes PRIDE (Paraphrase Robustness Index in Robotic Instructional DEviation), a metric that quantifies paraphrase difficulty using semantic and syntactic factors.method 90%
• Existing benchmarks for evaluating linguistic variation in VLA models have key limitations, such as treating paraphrasing as one axis among broader perturbations or conflating it with task-level semantic changes.limitation 85%
• VLA models in downstream robotic settings are typically fine-tuned with limited data, leading to overfitting to specific instruction formulations and poor robustness to paraphrased instructions.limitation 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card