SALA uses language descriptions of unpaired human and robot videos to estimate semantic similarity and establish cross-embodiment correspondences in a shared latent action space. The method requires no paired data and enables embodied foundation models to expand through both pretraining and post-training. Adding general and task-specific human videos can respectively reduce robot action prediction errors and improve real-world task success rates.
Source:Hugging Face 社区论文精选 · Read Original
Leave a Reply