Digital Cousin is a concept proposed by a team led by Professor Fei-Fei Li of Stanford University in 2024. It aims to provide a more efficient and economical solution for robot training. This concept has changed the way robots learn and has attracted widespread attention.ACDC: Automated Creation of Digital Cousins for Robust Policy Learning".

Rather than pursuing a one-to-one correspondence with the real object, the digital cousin focuses on similar geometric and semantic qualities, thereby generating practical training data at a lower cost. It is a virtual asset or scene that, unlike a digital twin, does not explicitly simulate its real-world counterpart, but still exhibits similar geometric and semantic functionality. This approach can simultaneously reduce the cost of real-to-simulation generation while increasing the generalizability of learning.

Digital cousins can be used for robot training by providing a virtual environment similar to the real world to train robot policies while reducing costs and improving cross-domain generalization capabilities. This approach generates a fully interactive simulation scene from a single RGB image and consists of 3 consecutive steps: information extraction, digital cousin matching, and scene generation.

Experimental results show that the robot strategy trained with digital cousins achieved a success rate of 90% in zero-shot virtual-to-real migration, far exceeding the 25% of digital twin training. This shows that digital cousins have relatively good performance both in-distribution and out-of-distribution, proving its advantage in generalization.