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.
RJB-Instructions-5 is the 90-scene RJB-Instructions set evaluated in the
paper (5 scenes per taxonomy case). RJB-Instructions is its 900-scene
expansion (50 per case) with regenerated images and instructions.
Columns: id (uuid, unique across the repo), dataset, category,
images, benign, malicious, target.
- Downloads last month
- 3,116