Autonomous Predicate Discovery for Neuro-Symbolic Planning

This direction is merged into another direction: predefined predicate dependency. Its last version stays readable and citable here.

Version 1 of 134 papers37 conceptsfrozen 2026-10-04

In the quest to build agents that can both perceive the world and reason about it, researchers have long faced a "representation gap." On one side, deep reinforcement learning (DRL) excels at processing raw, continuous sensory data like pixels or joint torques. On the other, symbolic AI excels at long-horizon planning and logical generalization.

The bridge between these two worlds is the symbolic predicate—a high-level concept like On(BlockA, BlockB) or Holding(Key). Historically, these predicates were hand-coded by human experts. However, a growing body of research is moving toward autonomous predicate discovery, where agents learn to "invent" the logical vocabulary they need to solve tasks directly from experience.

The Functional Role of Symbols

To understand why we want to learn symbols, we must first understand what a symbol is for a robot. In the framework established by George Konidaris and colleagues in Constructing Symbolic Representations for High-Level Planning, symbols are not just arbitrary labels; they are strictly determined by the agent's available actions (options).

A symbol represents a "precondition" (the state required to start an action) or an "effect" (what changes after the action). If an agent has a skill to "grasp a cup," the symbolic representation of the world must include a predicate that distinguishes between states where the cup is graspable and states where it is not. By grounding symbols in the feasibility of actions, agents can convert continuous environments into discrete PDDL (Planning Domain Definition Language) models, enabling planning speeds orders of magnitude faster than traditional search.

Bilevel Planning: Bridging Logic and Geometry

Once an agent has a symbolic vocabulary, it can perform bilevel planning. As described by Rohan Chitnis, Tom Silver, and Leslie Pack Kaelbling in their work on Neuro-Symbolic Relational Transition Models (NSRTs), this involves a high-level search for a sequence of logical operators (e.g., Pick(obj) -> Place(obj, tray)) followed by a low-level search for the precise continuous trajectories needed to execute those steps.

The power of this approach lies in its ability to generalize. Because the high-level logic is relational (it talks about "objects" rather than specific pixel coordinates), a robot trained on three blocks can plan for ten blocks without further training. However, as noted by Nishanth Kumar and others, these systems often hit a "manual bottleneck": they still rely on humans to provide the initial set of predicates.

Breaking the Manual Bottleneck

The most recent frontier in this field is the move toward predicate invention. Instead of being told what Clear(block) means, the agent must discover that "clearness" is a relevant property for its planning success.

Several strategies have emerged to solve this:

  1. Explanatory Invention: Quentin Delfosse and Hikaru Shindo introduced EXPIL (Explanatory Predicate Invention for Learning), which extracts predicates from an agent's experience. By identifying which features explain the success or failure of a policy, the agent "invents" the logical concepts necessary to describe its world.
  2. Neural Predicate Discovery: In Bilevel Learning for Bilevel Planning, Bowen Li proposes IVNTR, which interleaves symbolic search with neural network training. The system "invents" neural predicates—classifiers that act as symbolic atoms—directly from robot demonstrations, bridging the gap between raw sensor data and high-level logic.
  3. Active Exploration: To learn these symbols efficiently, agents cannot just wander randomly. Garrett Andersen and George Konidaris demonstrated in Active Exploration for Learning Symbolic Representations that agents can use uncertainty-aware models to target regions of the state space where their internal symbolic model is most likely to be wrong, drastically reducing the data needed to build a reliable world model.

Why It Works: The Efficiency of Abstraction

The intuition behind this research is that abstraction is a compression of the search space.

A standard DRL agent might need thousands of episodes to learn a task because it treats every pixel change as equally important. In contrast, neuro-symbolic frameworks like Detect, Understand, Act (DUA) by Ludovico Mitchener and Alessandra Russo can solve complex physical reasoning tasks with as few as seven examples (Detect, Understand, Act).

By focusing only on "necessary changes"—a concept explored by Nishanth Kumar in Learning Efficient Abstract Planning Models—the agent ignores irrelevant details (like the exact position of a background object) and focuses on the logical transitions that actually matter for the goal. This results in models that are not only more efficient but also human-interpretable, as the learned rules can be read and verified by people.

Papers in this version

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_autonomous_predicate_discovery_for_neuro_98fa7a1a,
  title = {Autonomous Predicate Discovery for Neuro-Symbolic Planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/autonomous-predicate-discovery-for-neuro-symbolic-planning/rdr_98fa7a1ad39549a08d9113668aaf0976},
  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_autonomous_predicate_discovery_for_neuro_98fa7a1a_v1,
  title = {Autonomous Predicate Discovery for Neuro-Symbolic Planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/autonomous-predicate-discovery-for-neuro-symbolic-planning/rdr_98fa7a1ad39549a08d9113668aaf0976/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}