Automating the Symbolic Bottleneck in Robot Planning

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

Version 1 of 134 papers34 conceptsfrozen 2026-10-04

For decades, robotics has been caught between two powerful but incompatible paradigms. On one side is symbolic planning, where robots use formal logic (like PDDL) to solve complex, multi-step tasks with mathematical rigor. On the other side is connectionist learning, where neural networks ingest raw sensor data to handle the messy variability of the real world.

The bridge between these two—the set of logical rules and "predicates" (like IsHolding(block) or At(robot, stove)) that tell a planner how the world works—has historically been a manual bottleneck. As A. Patricia Ambler and colleagues noted as early as 1973 in A Versatile Computer-Controlled Assembly System, relying on manual coordinate inputs and hand-coded instructions prevents robots from being versatile. This "symbolic bottleneck" requires human experts to painstakingly define every possible state for every new environment, a process that simply does not scale.

A new wave of research is now automating this process, allowing robots to "invent" their own symbolic logic directly from experience.

Inventing the Language of Action

To plan, a robot needs abstractions. It doesn't need to know the exact RGB value of every pixel; it needs to know if a drawer is Open or Closed. Traditionally, a human would write a piece of code to detect this.

In Predicate Invention for Bilevel Planning, Tom Silver and his team demonstrated that these abstractions can be discovered autonomously. By optimizing a surrogate objective that estimates how much a specific predicate simplifies the planning process, the system "invents" the symbols it needs to solve long-horizon tasks faster.

This concept is taken further by Naman Shah and Jayesh Nagpal in their work on LAMP. Their framework allows robots to derive relational concepts—the "grammar" of the task—from raw, unannotated video. This allows a robot trained on a simple task with two objects to "zero-shot" generalize to scenarios with eighteen objects, because it has learned the underlying logic rather than just a specific motion.

LLMs as Automated Domain Engineers

The rise of Large Language Models (LLMs) has provided a new shortcut. LLMs possess vast "common sense" about how the world works (e.g., you must open a bottle before pouring from it), but they struggle with the rigid execution required for physical robots.

Researchers are now using LLMs to act as "domain engineers." In PDDLLM, Jinbang Huang and Zhanguang Zhang show that an LLM can watch a single human demonstration and automatically generate a formal PDDL domain. This combines the semantic reasoning of the LLM with the physical grounding of a simulator, outperforming standard LLM-based planners by ensuring the generated plans are actually executable.

Similarly, Haoming Ye and Panpan Cai introduced UniDomain, which pre-trains a unified symbolic domain from real-world demonstrations. Instead of starting from scratch for every kitchen or factory, the robot starts with a foundational "logic" that it can adapt to specific environments.

Grounding Logic in Pixels

A symbol is useless if the robot can't "see" it in the real world. This is the problem of grounding.

VisualPredicator, developed by Yichao Liang and Nishanth Kumar, solves this by having the robot write its own perception code. The system uses Vision-Language Models (VLMs) to generate Python snippets that act as "Neuro-Symbolic Predicates." If the robot needs to know if a surface is "slippery"—a property not in its original code—it can generate a visual check for that specific attribute.

This iterative loop is also central to UniPred by Bowen Li and Tom Silver. By using neural network training loss to inform LLM hypotheses about which predicates are missing, the robot effectively "debugs" its own understanding of the world until its symbolic model matches physical reality.

The Self-Improving Robot

The ultimate goal of this research direction is a robot that requires no expert intervention to learn new skills. Frameworks like Self-CriTeach by Jinbang Huang et al. envision a self-improving loop: the robot attempts a task, uses an LLM to critique its own symbolic domain based on failure, and regenerates its logic.

Even when humans are involved, the interaction is shifting from "coding" to "coaching." InterPreT, proposed by Muzhi Han and Yuke Zhu, allows non-experts to provide natural language feedback (e.g., "You forgot to turn on the stove first"). The system then translates that feedback into PDDL operators, allowing the robot to learn the formal rules of the world through simple conversation.

By automating the creation of symbolic domains, these researchers are moving robotics away from rigid, hand-crafted scripts and toward systems that can reason, adapt, and define their own logic in an unpredictable world.

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_automating_the_symbolic_bottleneck_in_ro_7ab1fd85,
  title = {Automating the Symbolic Bottleneck in Robot Planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automating-the-symbolic-bottleneck-in-robot-planning/rdr_7ab1fd85a3504f0ba82de58460dbfd5f},
  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_automating_the_symbolic_bottleneck_in_ro_7ab1fd85_v1,
  title = {Automating the Symbolic Bottleneck in Robot Planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automating-the-symbolic-bottleneck-in-robot-planning/rdr_7ab1fd85a3504f0ba82de58460dbfd5f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}