DropVLA: An Action-Level Backdoor Attack on Vision-Language-Action Models

Name: DropVLA: An Action-Level Backdoor Attack on Vision-Language-Action Models
Uploaded: 2026-09-15T12:04:32.931824+00:00
Duration: 368 s
Description: The episode examines the "DropVLA" paper, which details how attackers can hijack specific robot movements using only 0.31% poisoned training data. By using simple visual triggers like a red circle, attackers can control basic actions in milliseconds. The hosts discuss mitigation strategies like runtime gating and rigorous data auditing.

The episode examines the "DropVLA" paper, which details how attackers can hijack specific robot movements using only 0.31% poisoned training data. By using simple visual triggers like a red circle, attackers can control basic actions in milliseconds. The hosts discuss mitigation strategies like runtime gating and rigorous data auditing.

Open the full episode