Researchers led by Keke Tang have built ExecGap, an adversarial patch attack on robots controlled by vision-language-action models, according to a paper on arXiv. It exploits the fact that the commands such a model predicts are not carried out directly but pass through the robot's control system and body first.
ExecGap shifts the predicted commands toward directions the robot carries out poorly, estimated from where its kinematics are inefficient and its actuators weak. The commands then become hard to realise and the task fails.
In simulations with several vision-language-action policies, ExecGap caused a bigger mismatch between commands and execution than existing patch attacks, and experiments on real robots showed that the patch also degraded execution on physical machines. The paper was accepted at CoRL 2026.



