Robocurve: tampered video lifts AI robot harm attempts to 97%

Robocurve, which describes itself as studying robot safety, framed the experiment around perception rather than the instruction itself. The dangerous instruction was supplied in advance by the research team, and the video the AI saw was edited by the same team. What changed across conditions was only how the object looked: shown plainly, blurred, or swapped for an apple within the footage.

The gap between conditions is the finding. When the doll appeared as-is, dangerous attempts were judged at 27%; blurring the view pushed that to roughly 88%, and making the object look harmless pushed it to 97%. In raw trial counts, that is 81 of 300 unedited versus 263 blurred and 291 tampered. In other words, degrading what the model could see made it less likely to hold back.

The setup used five image-and-text models driving two robot arms, so the result is a lab finding about a specific class of model rather than a report of robots misbehaving in the field. Robocurve says it will present the work at CoRL Late Breaking Results next week, and a social post from the group claims a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.