MIT proposes UQ-DAAAM to correct mislabeled robot memory

MIT researchers propose the UQ-DAAAM system to identify and correct potentially erroneous object descriptions in robot memory. The system adds uncertainty quantification to the DAAAM framework, quantifying uncertainty by comparing multiple descriptions of the same object from different viewpoints, then selecting additional viewpoints that reduce uncertainty, and using a large language model to fuse the multiple descriptions before writing them back to memory, without rerunning the vision-language model. According to the research, this correction step can disambiguate about 92% of objects in the robot's environment, requiring only two additional descriptions per object; object description accuracy improves by 5% over the current state-of-the-art method. The research will be presented at the Conference on Neural Information Processing Systems (NeurIPS).

AI 根据报道生成 · 5 days ago更新