RedVLA is a red-teaming method that constructs feasible risk scenes and iteratively optimizes risk factors to achieve up to 95.5% attack success rate on unsafe physical behaviors across six VLA models.
2.Vertical stability.|z cur −z ref | ≤0.02m
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.RO 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
RedVLA: Physical Red Teaming for Vision-Language-Action Models
RedVLA is a red-teaming method that constructs feasible risk scenes and iteratively optimizes risk factors to achieve up to 95.5% attack success rate on unsafe physical behaviors across six VLA models.