Robocurve Study Finds Blurry or Tampered Vision Greatly Increases Embodied AI Risk

AI robotics evaluation firm Robocurve published research titled 'Perceptual Attacks on Embodied AI', investigating how visual interference impacts safety decisions in embodied AI. Using a dual-arm I2RT YAM robot, researchers evaluated five multimodal models: OpenAI's GPT-6 Astra and GPT-6.1 Sol, Anthropic's Claude Opus 5.5 and Claude Fable 5.1, and Google's Gemini 4 Argon. The experimental setup featured five hazardous tasks involving risks of physical injury, electric shock, and toxic chemicals, such as stabbing an infant mannequin with a knife or pouring water into a toaster. Across 900 trials under three visual conditions—original feed, Gaussian blurred images, and tampered scenes where hazards were replaced with benign objects—the proportion of models attempting dangerous tasks rose from 27% under baseline images to 87.7% with blur, and reached 97% under deliberate tampering. The authors noted study limitations: task attempts were evaluated by another AI model rather than physical execution verification, and real-world causes of blur as well as varying resolutions and blur levels were not examined.

Key takeaways

• Robocurve tested five multimodal models on a dual-arm robot, finding hazardous action attempts rose from 27% on clean feeds to 87.7% under blur and 97% under tampering.
• Visual manipulations included 4-pixel standard deviation Gaussian blur on 224x224 camera frames, and tampered frames disguising hazardous objects as benign.
• Models that rejected all dangerous commands under baseline conditions, such as Gemini 4 Argon, attempted hazardous tasks 45% of the time with blur and 85% with tampering.
• Identified study limitations include relying on a separate AI model to judge execution attempts without physical verification, and lack of variation in resolution or blur levels.