When we think about classic examples of structured symbolic planning, robotics examples come first to mind, like a robot planning how to make a cup of tea. Instead of learning from data, it reasons over explicit symbols, rules, and goals. But use cases can extend into other areas like banking, finance and supply chain planning as well, in tandem with a portfolio of agents tethered in the solution.

Logistics is a classic and highly intuitive example of a structured symbolic planning scenario, frequently used in automated planning and PDDL (Planning Domain Definition Language) benchmarks.

​For example, there can be a scenario where the objective is to move packages from their initial locations to specific destinations using a fleet of trucks and airplanes.

​Here is how a symbolic planning system structures this problem using Objects, Predicates (States), Actions, and Goals:

​1. The Objects (The Domain)

​First, we define the physical entities involved in the world:

​Locations: In_Office_A, In_Office_B, Airport_1, Airport_2
​Vehicles: Truck_1, Airplane_1
​Cargo: Package_X

​2. The Current State (Initial State)

​The system represents the exact state of the world using logical predicates. Before any actions are taken, the world looks like this:

​At(Package_X, In_Office_A) (The package is currently at Office A)
​At(Truck_1, In_Office_A) (Truck 1 is parked at Office A)
​At(Airplane_1, Airport_1) (The airplane is at Airport 1)

​3. The Goal State

​This is the target condition that the planner must satisfy:
​At(Package_X, In_Office_B) (Package X must end up at Office B)

​4. Available Actions & Rules

​Every action has strict symbolic Preconditions (what must be true to execute it) and Effects (how the world changes after execution).

​Action: Load_Truck(package, truck, location)
​Preconditions: At(package, location) AND At(truck, location)
​Effects: In(package, truck) AND NOT At(package, location)
​Action: Drive_Truck(truck, from_location, to_location)
​Preconditions: At(truck, from_location)
​Effects: At(truck, to_location) AND NOT At(truck, from_location)
​
​If an LLM lacks structured planning training, it might hallucinate a step like “Fly the truck to Airport 2” or try to “Load Package X into the airplane” while the airplane is still at a completely different airport.

This paper aims to address this limitation.

​To solve such limitations, the authors introduce PDDL-INSTRUCT, a novel instruction tuning framework designed to explicitly teach LLMs the underlying formal structures of automated planning.

​Instead of just prompting models or fine-tuning them on simple text outcomes, the framework breaks down planning verification into explicit, atomic reasoning chains.

Gist is that a symbolic planning framework (like the PDDL-Instruct framework mentioned previously) forces the model to mathematically verify that every precondition is satisfied before it can declare a plan valid.

#AI #ArtificialIntelligence #RoboticsAI #Robotics