Tyr
Tyr is a planner in C++20 and Python with ground and lifted configurations on a shared core. Lifted planning avoids grounding the task in full, which helps when grounding itself is the bottleneck.
- Author
- Dominik Drexler
- Institution
- Linköping University
- Kind of planning
- Classical and numeric
- Source
- github.com/planning-and-learning/tyr (GPL-3.0); on PyPI as
pytyr
How it works
Tyr's optimal configurations use A* with a blind heuristic. Its agile and satisficing configurations use lazy greedy best-first search, with a goal-count heuristic when ground and a lifted additive heuristic when lifted. In its lifted configurations, per its abstract, "successor generation avoids full grounding by deriving candidate action bindings from substitution-consistency graphs and Datalog facts".
What it handles
Classical PDDL with ADL features, derived predicates and action costs, and numeric conditions and effects. It does not handle durative actions. See the feature comparison.
In the competitions
At IPC 2026, Tyr's ground configuration won the overall optimal track, and it took a medal in the overall agile track, according to the published results.
Results in our benchmarks
| Track | Problems | Tyr | Autoplan |
|---|---|---|---|
| IPC 1998–2018 classical | 2,742 | 1,111 | 2,450 |
| IPC 2023 classical | 140 | 9 | 109 |
| IPC 2023 numeric | 400 | 109 | 383 |
| IPC 2026 numeric | 320 | 66 | 270 |
Tyr 0.0.15: problems solved with a plan the VAL validator accepts, within 60 seconds and 4 GB of memory on one core. These are our own runs, not competition results, and a planner written for other goals or for a different kind of problem can score low here while being the right tool for its purpose. How we measured.
Our runs used version 0.0.15, the one entered in the competition. Tyr has moved on since.
Papers
- Dominik Drexler. Tyr: A Ground and Lifted Numeric Planner for the IPC 2026. Planner abstract, IPC 2026 numeric track.
- Dominik Drexler. Lifted Successor Generation in Numeric Planning. arXiv:2511.00673, 2025.