Tracks

A track is a combination of two choices: what kind of input a planner has to read, and what its plan has to achieve. Not every combination is offered – the table below shows what we will run.

The Tracks

Note that here we only point out what the input will be and what we demand for valid outputs – this still abstracts away from the exact scoring function used to determine the winner. This will be mentioned deeper down.

• Optimal tracks: Here, we demand that any solution returned is optimal. Returning a non-optimal solution means disqualification in that domain; a non-optimal solution in a second domain means disqualification from all domains (see plan validity and disqualification). In this case, two cost models are supported: none (equals unit costs) and general action costs (see below). Note that the scoring function used will be a combination of coverage and time needed.
• Non-optimal tracks: These tracks won’t use any cost model. However, note one important difference:
  • In the two tracks (total-order and partial-order) that are not about makespan, plan quality doesn’t matter at all, so scoring is purely done on coverage and time needed. Output plans are sequential as in the optimal tracks.
  • In the makespan track (partial-order only), plans with shorter makespan (under implicit unit-durations, i.e., critical path length) are preferred. Thus, this is the only track where the output supports being a partially ordered plan. Plan quality will be taken into account, but so will the planner’s runtime. This still differs from satisficing tracks from earlier years where only plan quality counts, as there one could use the full time limit to find better plans with no downside. Here, it might still be beneficial to return the first plan found (even if not optimal) since the planner’s runtime goes into the score.
• Unsolvability track: Here, planners have to recognize unsolvable problems. For each problem, a planner returns either a plan, as in all other tracks, or the string unsolvable if it proved that no plan exists. Since we do not ask for a certificate of unsolvability, an unsolvable answer cannot be verified on its own. We therefore mix the unsolvable problems with hard solvable ones, so that unsolvable cannot simply be guessed: a wrong unsolvable is treated like an invalid plan and leads to disqualification (see plan validity and disqualification). All problems are totally ordered, and no cost model is used.

Whenever a planner returns a plan, it also returns the (ground) methods used to obtain it, i.e., the decomposition hierarchy; see output format.

How Configurations Are Scored

A planner may enter a track with up to two configurations – variants that differ, e.g., in the heuristic or search strategy used (see planners and configurations). In each track, every configuration gets a single score that combines coverage with runtime – how many problems it solved, and how quickly. There is no separate agile track and no separate coverage track; both aspects enter the same number. The exact function is still being settled and will be published soon. How these scores determine the places is described under final ranking.

Whatever that function turns out to be, we will publish the complete raw data once the competition is over, ordered by different possible ranking functions. Anyone who prefers a different ranking – coverage alone, runtime alone, or some combination of those, different from what we chose – can compute it themselves, and may well arrive at a different winner.

Input Language and Benchmarks

All problems are given in HDDL. Which HDDL features are used, and where the benchmark problems come from, is described under benchmarks.

Last updated 6 October 2026.