Neural Network Learns Action Schemas from Unlabeled Traces
A new neural network architecture learns lifted action schemas for classical planning from fully observed states but unobserved action arguments. The work, published on arXiv (2605.13282), addresses a simplification of the problem of learning planning domains from sequences of images and action labels. The approach aims to achieve near-perfect performance on this task, which is a step toward structural generalization in AI planning.
Key facts
- arXiv paper 2605.13282
- New neural network architecture for learning action schemas
- States are fully observed, action arguments are unobserved
- Simplification of learning from images and action labels
- Aims for near-perfect performance
- Addresses structural generalization in classical planning
Entities
Institutions
- arXiv