Differentiable Learning of Lifted Action Schemas for Classical Planning

2026-05-15 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new neural network architecture has been developed for learning lifted action schemas in classical planning domains, as detailed in arXiv:2605.13282, submitted on May 13, 2026. This architecture addresses the challenge of learning action schemas and simultaneously identifying action arguments from observed state changes, specifically when states are fully observed but action arguments are not. The work simplifies the broader problem of learning planning domains from image sequences and action labels, aiming for near-perfect solutions in this specific context. The resulting differentiable component is designed for integration into larger neuro-symbolic models. Evaluation across various planning domains demonstrates its ability to recover ground-truth structure, with additional experiments assessing robustness to observation noise and exploring a variation related to slot-based dynamics models.

Key takeaway

For research scientists developing neuro-symbolic AI, this work provides a robust differentiable component for learning planning domain dynamics. You should consider integrating this architecture to improve the accuracy of action schema learning, especially in scenarios where action arguments are unobserved but state changes are fully visible, thereby enhancing the generalization capabilities of your planning systems.

Key insights

A novel neural network learns lifted action schemas from state traces with unobserved action arguments.

Principles

• Classical planners excel with STRIPS/PDDL representations.
• Lifted schemas enable structural generalization.
• Learning schemas from data is a central challenge.

Method

The approach involves a neural network architecture that learns action schemas and identifies action arguments concurrently from fully observed states with unobserved action arguments, yielding a robust differentiable component.

In practice

• Integrate into larger neuro-symbolic models.
• Evaluate robustness to observation noise.
• Explore slot-based dynamics model variations.

Topics

• Differentiable Learning
• Lifted Action Schemas
• Classical Planning
• Neural Network Architecture
• Neuro-Symbolic Models

Best for: Research Scientist, AI Scientist

Related on AIssential

• Differentiable Learning of Lifted Action Schemas for Classical Planning — A neural network learns lifted action schemas and arguments from observed state changes with unobserved actions.
• Scaling Neuro-symbolic Problem Solving: Solver-Free Learning of Constraints and Objectives — A neuro-symbolic architecture with a probabilistic loss efficiently learns to solve NP-hard problems from natural inputs.
• Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability — Reasoning SFT can generalize cross-domain, but it is conditional on optimization, data quality, and model capability.
• Improving Vision-Language-Action Model Fine-Tuning with Structured Stage and Keyframe Supervision — Structured supervision significantly enhances Vision-Language-Action model fine-tuning for robotic manipulation.
• Beyond KV-Cache: Test-Time Training & Invariant Latent Topologies for ICL — New methods enable LLMs to learn during inference and generalize across diverse domains by modifying internal weight structures or identifying invariant neural subspaces.
• PLAN-S: Bridging Planning with Latent Style Dynamics for Autonomous Driving World Models — PLAN-S improves autonomous driving LWMs by using style-conditioned semantic cost maps for explicit risk and style modulation in planning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.