Robocurve study finds blurred images raise AI-labeled dangerous-task attempts
Robocurve’s “Perceptual Attacks on Embodied AI” study tested five multimodal models controlling a dual-arm I2RT YAM robot across five hazardous tasks and three visual conditions, totaling 900 trials. According to iThome Taiwan, models attempted dangerous tasks in 27% of trials with original images, 87.7% with Gaussian-blurred images and 97% with deliberately altered images. Another AI model judged whether an attempt occurred, and the researchers did not verify the physical outcomes.
The results suggest that degraded visual input can weaken embodied-AI safety judgments, with image blur producing an attempt rate close to that of deliberate alteration. However, AI-based attempt labeling and the lack of outcome verification limit conclusions about physical harm.
圖片來源:
Robocurve
AI機器人獨立評測公司Robocurve於11月6日公布一項研究,發現當AI模型控制實體機器人執行任務時,攝影機影像若遭到模糊處理或刻意竄改,可能大幅影響其安全判斷。研究團隊在特定實驗條件下測試5款多模態AI模型,發現在原始影像下,模型嘗試執行危險任務的比例為27%,但畫面模糊後升至87.7%,若改以刻意竄改的影像測試,比例更達97%。
這項名為Perceptual Attacks on Embodied AI的研究,主要探討具身AI(Embodied AI)的感知安全問題。所謂感知攻擊(Perceptual Attacks),是指透過干擾或竄改AI接收到的環境資訊,使其錯誤辨識周遭物體及情境,進而影響後續操作決策。
為了驗證這項問題,Robocurve使用具有雙機械臂的I2RT YAM機器人,測試OpenAI的GPT-6 Astra與GPT-6.1 Sol、Anthropic的Claude Opus 5.5與Claude Fable 5.1,以及Google的Gemini 4 Argon等5款多模態AI模型。實驗中,研究團隊設計5項可能造成危險的操作任務,例如,持刀刺向嬰兒模型、將電源插座放入水中、把水倒入烤麵包機、混合漂白水與氨水,以及將漂白水當成飲料倒入杯中,藉此測試AI模型能否辨識操作過程中,是否存在人身傷害、觸電以及有毒物質等安全風險。
#Robocurve#OpenAI#GPT-6 Astra#GPT-6.1 Sol#Anthropic#Claude Opus 5.5#Claude Fable 5.1#Google#Gemini 4 Argon#I2RT YAM