Towards Numerical TOHTN Planning with SMT-based HTN... | AI Research

Towards Numerical TOHTN Planning with SMT-based HTN-SAT Encoding
Hierarchical Task Network (HTN) planning is a powerful method for solving complex problems by breaking them down into smaller, manageable tasks. However, most existing HTN planners struggle to handle "numerical fluents"—variables like fuel levels, costs, or quantities—which are essential for real-world applications like robotics and logistics. This paper addresses this gap by extending standard SAT-based HTN planning to include numerical reasoning through Satisfiability Modulo Theories (SMT), creating a more expressive and capable planning framework. The same reasoning question is explored in GRAIN, which adds a research perspective.

Bridging Logic and Numbers

The researchers build upon existing SAT-based planning techniques, which represent the decomposition of tasks as a tree structure. While traditional SAT solvers are excellent at handling logical propositions (true/false states), they cannot natively perform arithmetic. By upgrading the underlying encoding to SMT, the authors allow the planner to treat numerical constraints as part of the logical problem. This means the planner can now verify if a sequence of actions is valid not just logically, but also numerically—for example, ensuring a robot has enough fuel to complete a series of tasks.

A New Benchmark for Evaluation

A significant challenge in this field has been the lack of a common standard for testing numerical HTN planners. To solve this, the authors introduced a new suite of seven numerical benchmark families. These include scenarios like "Transport-Fuel," which requires managing fuel consumption during deliveries, and "Transvasement," which involves complex water-jug-style liquid transfers. This benchmark suite provides a necessary foundation for the research community to compare different planning approaches on a level playing field. The same ai evaluation question is explored in Discriminative World Models for Web Agents, which adds a research perspective.

Performance and Robustness

The authors tested their new SMT-based planner, SibylSmt, against two existing numerical HTN planners: Siadex and Aries. The experimental results indicate that the SMT-based approach is highly competitive. SibylSmt demonstrated higher overall coverage and better "agile scores"—a metric measuring how quickly a planner finds a solution—across the new benchmark suite. The findings suggest that this SMT-based encoding is more robust, particularly in domains with complex, recursive structures where other planners often struggle to find a valid decomposition.

Future Directions

While the results are promising, the authors note that none of the tested planners were able to solve every instance in the new benchmark suite. This highlights that there is still significant room for improvement in numerical hierarchical planning. By providing a clear, effective baseline and a new set of benchmarks, this work aims to encourage further research into more expressive and efficient HTN planning models that can handle the complexities of real-world environments. The robotics story also surfaces in NVIDIA Launches Cosmos 3 Edge for..., adding another angle. as detailed in the full paper on Arxiv