Runtime planning—such as robots replanning while acting—requires a planner that can ingest user‑written PDDL (with durative actions, numeric resources, time windows, continuous change), respond quickly, and output only validated plans. Existing specialized planners excel on individual PDDL fragments but differ in coverage, latency, and the meaning of their failures.
Kairos addresses this by selecting heterogeneous search engines based on the semantics of each task, aiming for broad PDDL coverage from STRIPS to PDDL+ while staying competitive with specialized planners and making each answer’s guarantees explicit. The system first profiles a task, then chooses among five engines: ground heuristic search, compressed snap‑action temporal search, time‑step search for hybrid dynamics, lifted search, and other dedicated searches. Every candidate plan is checked by a validator independent of the planner’s internal transformations, and “unsolvable” is reported only when an exhaustive search provides a proof of impossibility.
On 1,756 tasks from nine comparative suites, Kairos solves more tasks under full limits than the strongest evaluated planner in each suite, with the largest gaps on numeric, temporal, PDDL3, and PDDL+ problems; it also leads at strict 100 ms and 1 s budgets. Resampling across domains shows a positive gain in six of the nine suites, while on classical planning against LAMA the difference is within noise. An additional 450 feature‑oriented tasks and 26 single‑feature probes demonstrate that Kairos handles all seventeen tested features, whereas no other evaluated planner covers more than eight. Ablation studies identify compressed temporal search and the novelty fallback as major contributors; a comparison with VAL attributes every disagreement to a PDDL+ convention or a documented VAL defect. Both the planner and the benchmark suite are open source.
Review