AI Research Proposes Framework for Learning Action Models from Visual Traces Without Supervision
A recent study presents a deep learning framework aimed at deriving lifted action models from sequences of state images without the need for action observation. This method concurrently learns state predictions, action predictions, and a lifted action model. To mitigate issues like prediction collapse and self-reinforcing mistakes, the authors introduce a mixed-integer linear program (MILP) that maintains logical consistency among predicted states, actions, and the action model across a selection of traces. Further learning is directed by pseudo-labels obtained from the MILP results. This research tackles the challenge of efficiently building models that reflect action preconditions and effects, essential for implementing AI planning in practical scenarios. The paper, titled "Learning Lifted Action Models from Unsupervised Visual Traces," can be found on arXiv with the identifier arXiv:2604.19043v1. The announcement is categorized as new, and its abstract details the methodology and its importance for enhancing AI planning capabilities.
Key facts
- The paper proposes a deep learning framework for learning lifted action models from state image sequences without action observation.
- It jointly learns state prediction, action prediction, and a lifted action model.
- A mixed-integer linear program (MILP) is introduced to prevent prediction collapse and self-reinforcing errors.
- The MILP ensures logical consistency across predicted states, actions, and the action model.
- Pseudo-labels from the MILP solution guide further learning.
- The research addresses efficient construction of models for AI planning in real-world domains.
- Prior work has explored learning from high-level descriptions of state and/or action sequences.
- The paper is available on arXiv under identifier arXiv:2604.19043v1 with an announcement type of new.
Entities
Institutions
- arXiv