Automated Induction of Symbolic Planning Representations

This direction is merged into another direction: manual domain modeling. Its last version stays readable and citable here.

Version 1 of 195 papers99 conceptsfrozen 2026-10-04

Classical AI planning is remarkably powerful: if you provide a computer with a formal description of the world—what objects exist, what actions can be taken, and what the rules of cause-and-effect are—it can find a guaranteed path to a goal. However, this power comes with a steep price known as the knowledge engineering bottleneck. For decades, these formal descriptions (often written in languages like PDDL) had to be painstakingly hand-coded by human experts.

A growing body of research seeks to eliminate this manual step by enabling agents to autonomously "learn the rules of the game." By observing state transitions, analyzing image sequences, or interpreting natural language, these systems induce the symbolic models they need to reason.

The Knowledge Engineering Bottleneck

In traditional symbolic AI, a planner is like a world-class navigator who has never seen a map. If a human draws the map perfectly, the navigator is unstoppable. But in the real world, maps are rarely provided. As noted by Masataro Asai and Alex Fukunaga in Classical Planning in Deep Latent Space, the requirement for human experts to manually encode environment logic limits the scalability and autonomy of agents in unfamiliar domains.

Early attempts to solve this, such as CHEF by Kristian J. Hammond, focused on case-based reasoning—modifying past successful plans to fit new problems. While this reduced the need to plan from scratch, it still relied on hand-coded causal models to understand why a plan failed. The modern challenge is to move entirely beyond these pre-defined models.

Bridging the Subsymbolic-Symbolic Gap

The most difficult version of this problem is learning symbols directly from "subsymbolic" data, such as raw pixels. How does a robot move from seeing a grid of RGB values to understanding the logical concept of (holding block-A)?

The LatPlan framework, introduced by Asai and Fukunaga, uses Variational Autoencoders (VAEs) to map images into a discrete latent space. This allows the system to treat clusters of visual data as symbolic states, enabling a classical planner to solve puzzles like the 8-puzzle directly from images.

More recently, researchers have moved toward "lifted" models—rules that apply to categories of objects rather than specific instances. Kai Xi developed ROSAME-I, which uses a neuro-symbolic approach to learn human-readable PDDL models from visual traces. By using differentiable relaxations of logic, the system can "backpropagate" through the rules of physics to find the simplest symbolic description that explains the observed movement of objects in a video.

Learning from Interaction and Traces

If an agent cannot see the world, it can still learn by observing "traces"—sequences of actions and their outcomes. The goal here is Action Model Acquisition: inferring the preconditions (what must be true to act) and effects (what changes after acting) of every possible move.

  • Safe Learning: Brendan Juba and Roni Stern proposed SAM Learning, which focuses on "conservative" induction. It only adds a rule to its model when it is absolutely certain, ensuring that any plan it generates is guaranteed to succeed.
  • Handling Noise: Real-world data is rarely clean. Hankz Hankui Zhuo has explored learning models from disordered, noisy, or even crowdsourced data. In CAMA, the system combines noisy "wisdom of the crowd" suggestions with limited real-world observations to build a functional planning model.

Leveraging Language and Common Sense

The latest frontier in this research involves using Large Language Models (LLMs) as a source of "common sense" to jumpstart the modeling process. While LLMs are prone to hallucination and struggle with complex logic, they are excellent at proposing plausible rules.

The Ada (Action Domain Acquisition) framework by Jiayuan Mao and Joshua B. Tenenbaum uses LLMs to propose symbolic operators which are then rigorously verified through environment interaction. Similarly, PSALM by Robin Jia allows agents to infer action semantics by combining LLM reasoning with environmental feedback.

By using Vision-Language Models (VLMs), systems like pix2pred can even "invent" new predicates—discovering concepts like "fragile" or "heavy"—simply by watching a few demonstration videos. This represents a full circle in the research: the system starts with raw pixels, uses the broad knowledge of a foundation model to suggest symbols, and then refines those symbols into a precise, formal model that a classical planner can use to solve novel, long-horizon tasks.

Papers in this version

Showing 60 of 95 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_automated_induction_of_symbolic_planning_879994d1,
  title = {Automated Induction of Symbolic Planning Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-induction-of-symbolic-planning-representations/rdr_879994d1f62b4a1490bb95a6f54ecf88},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_automated_induction_of_symbolic_planning_879994d1_v1,
  title = {Automated Induction of Symbolic Planning Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-induction-of-symbolic-planning-representations/rdr_879994d1f62b4a1490bb95a6f54ecf88/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}