Researchers have developed a new deep learning framework for learning lifted action models from sequences of state images without direct action observation. This method jointly learns state prediction, action prediction, and a lifted action model. To address prediction collapse and self-reinforcing errors, a mixed-integer linear program (MILP) is introduced to find logically consistent solutions, which then provide pseudo-labels to guide further training and improve convergence. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
RANK_REASON This is a research paper detailing a new deep learning framework and methodology for learning action models from visual data.