up-kairos

Kairos, a temporal and numeric PDDL planner, as an engine of the Unified Planning Framework (UPF).

Kairos plans in the same process (no files, no subprocess), anytime, and proves when no plan exists. Every plan is validated before it is returned.

pip install up-kairos

import up_kairos
from unified_planning.shortcuts import OneshotPlanner

up_kairos.register()  # until UPF lists Kairos among its engines

with OneshotPlanner(name="kairos") as planner:
    result = planner.solve(problem, timeout=10)
    print(result.status, result.plan)

Operation modes

Results:

• SOLVED_OPTIMALLY when Kairos proves the plan optimal for the problem's metric.
• UNSOLVABLE_PROVEN when it proves there is no plan: the search space runs out, the timed literals take what the goal needs too soon, or there is no plan even without durations.
• UNSOLVABLE_INCOMPLETELY when the engines that apply ran out of states without being complete for the task.
• TIMEOUT, MEMOUT, UNSUPPORTED_PROBLEM.

Kairos' diagnostics, including why there is no plan, are in result.log_messages. Its statistics are in result.metrics: expanded nodes, states, plans and peak memory.

Problems

Kairos plans with:

• Classical planning, with negative, disjunctive, existential and universal conditions and conditional effects.
• Numeric planning, linear or not, with bounded numeric fluents and undefined values.
• Object fluents.
• Temporal planning, with required concurrency, timed effects (timed initial literals and fluents), duration inequalities and durations that read fluents.
• Continuous effects, processes and events (PDDL+).
• Trajectory constraints and state invariants.
• Quality metrics: action costs, plan length, makespan, final value, and oversubscription.

It plans 52 of the 66 example problems of UPF 1.3. The rest are hierarchical, scheduling, integer or boolean parameters, intermediate conditions, timed goals and undefined booleans.

Notes:

• Writing for Kairos. Some things go beyond what UPF's PDDL writer writes, so this engine writes them itself:
  • object fluents, as PDDL 3.1;
  • oversubscription, as goal preferences and a metric of the gain lost;
  • bounded fluents, as state invariants.
• One instance at a time. A durative action does not overlap itself, as in POPF: SELF_OVERLAPPING is not declared.
• Simultaneous happenings follow PDDL 2.1 and VAL. Happenings that interfere need epsilon between them (problem.epsilon, 0.001 by default). Kairos' plans keep that separation.
• PDDL+ plans and their end. A PDDL+ plan ends when the goal is checked, after letting processes and events act. UPF's plans have no end time, so it comes in result.metrics["kairos_plan_end"], and PlanValidator answers UNKNOWN for PDDL+.

Options

OneshotPlanner(name="kairos", params={...}):

• anytime
• timeout: the default when the call gives none.
• threads
• memory_limit_mb
• engine: "auto", "m1", "m2" or "m3".
• search
• seed
• stop_after_no_improvement
• target_cost

License

Apache-2.0.