MIT's UQ-DAAAM Helps Robots Revise Uncertain Memories
MIT researchers introduced UQ-DAAAM, a robot memory system that estimates uncertainty in object descriptions, selects clearer views from existing footage and updates stored descriptions. In testing on the OC-NaVQA benchmark, descriptive-question accuracy increased from 71.1% for DAAAM to 76.1% for UQ-DAAAM. With at most two additional viewpoints, uncertainty scores for 92.3% of selected objects fell below a preset threshold. A lower uncertainty score does not necessarily mean the revised description is correct: even after receiving clearer views, the system once misidentified a bicycle rack as a telephone booth.
The researchers say reliable environmental memory is important for robots performing complex, long-running tasks in homes, hospitals and factories. They also describe the ability to recognize uncertain information as an important step toward safer application.
机器人在路上行走时,看到屋顶上有一个灰色物体。由于距离较远、画面不够清晰,它最初没能准确描述这个物体,随后又将它认成了一只松鼠。但随着机器人继续移动,摄像头从新的角度拍下了这个物体,系统最终发现,它其实是一根烟囱。
那么,在获得更多信息后,机器人能不能像人一样,回头修正之前的判断?
近日,麻省理工学院的研究人员介绍了一套新的机器人记忆系统 UQ-DAAAM,试图解决这个问题。它能够检查记忆中哪些物体描述可能不可靠,再从已有画面中寻找更清晰的视角,补充信息并更新记忆。
在测试中,对于被标记为不确定的物体,系统最多补充两个观察视角,就使 92.3% 的物体描述的不确定性降到了设定阈值以下。在一项时空问答测试中,与原有系统相比,它回答描述性问题的准确率从 71.1% 提高至 76.1%,对物体位置和出现时间的判断也有所改善。