RoboBench
Benchmark artifacts for jailbreak attacks on vision-language robot planners.
One config per artifact; every artifact is self-contained (images included).
Vanilla artifacts carry the original malicious and benign instruction per
frame; <attack>_<base> artifacts carry the attack's rewritten malicious
instruction with the same frames and benign instruction.
Columns: id (uuid, unique across the repo), dataset, category,
images, benign, malicious, target.
- Downloads last month
- 290