Planning with Temporal Memory: A Polynomial Trick for Complex Goals
This presentation explains how planning problems with temporally extended goals—goals that depend on execution history—can be efficiently solved without building expensive automata. By converting Pure-Past Linear Temporal Logic goals into standard reachability problems through clever bookkeeping, the approach enables off-the-shelf planners to handle complex temporal specifications with minimal overhead.Script
Most planners are designed to reach a single goal state, but what if your goal depends on the entire journey? This paper shows how to plan when success is defined not by where you are, but by what you've done along the way.
Pure-Past Linear Temporal Logic lets you specify goals like achieve task A, then B, then C, or ensure a safety condition has always held. These temporal constraints are non-Markovian because whether you've satisfied the goal depends on history, not just your current state.
The core insight is deceptively simple. Instead of building an automaton for your temporal goal, augment each planning state with bookkeeping fluents that remember just enough about the past. These fluents store the truth of relevant temporal subformulas from the previous step, creating a compact summary of history.
Every action in the compiled problem carries the same temporal update mechanism. After executing its normal effects, the action updates the bookkeeping fluents based on which temporal subformulas are currently true. The temporal goal becomes a standard reachability condition: reach a state where the derived predicate for your formula holds.
The experiments demonstrate that compilation overhead is minimal. Runtime on compiled problems closely tracks runtime on original reachability problems, and the approach scales better than competing temporal-goal compilers, especially as goal complexity increases.
By reducing temporally extended goals to standard reachability, this work lets existing high-performance planners tackle complex temporal specifications without redesign. To explore more research like this and create your own video summaries, visit EmergentMind.com.
