Adversarial text printed in a robot's workspace hijacks VLM planners in 5 to 29 percent of trials, with simple defenses reducing failures to near zero.
Advanced Robotics , volume=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.RO 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Hijacking Robots with a Piece of Paper: A Systematic Study of Physical Prompt Injection in VLM-Controlled Robots
Adversarial text printed in a robot's workspace hijacks VLM planners in 5 to 29 percent of trials, with simple defenses reducing failures to near zero.