Temporal Planning

Unlike classical planning which treats actions as instantaneous, temporal planning models actions with durations and events with timestamps. The planner must schedule actions on a continuous or discrete timeline, ensuring that temporal ordering constraints (e.g., Action B must start at least 5 seconds after Action A finishes) are satisfied. This is fundamental for robotics, where movement and sensing operations have non-zero execution times.

A core capability is reasoning about parallelism. Temporal planners can schedule actions to overlap in time if they do not conflict in terms of resources or preconditions. This is critical for efficiency in real-world systems. For example:

• A robot arm can be moving to a location while its camera is scanning for an object.
• Two independent agents can perform tasks simultaneously. The planner must manage potential interactions and ensure non-interference between concurrent actions.

Plans must satisfy complex temporal constraints beyond simple sequences. These include:

• Metric Constraints: Specifying minimum or maximum delays between events (e.g., end(A) + 5 <= start(B)).
• Deadlines: A goal must be achieved within a specific absolute time.
• Windows of Opportunity: An action can only be executed during a specific time interval. These constraints are essential for tasks like industrial process control, where timing is critical for synchronization and safety.

Temporal planning must account for the consumption and production of continuous resources over time, such as battery charge, fuel, or the level of a liquid in a tank. Actions are modeled with resource profiles that specify their usage rate throughout their duration. The planner ensures that resource levels never fall below a minimum (avoiding depletion) or exceed a maximum (avoiding overflow) at any point in the planned timeline.

In physical systems, action durations and outcomes are often uncertain. Advanced temporal planning incorporates temporal flexibility and contingency planning. Instead of producing a single rigid schedule, planners may generate:

• Temporal Networks with Uncertainty (STNUs): Plans where some timepoints are controlled by the agent and others are observed from the environment.
• Flexible Time Windows: Specifying ranges for when actions can start, allowing for online adjustment during execution to accommodate delays.

In robotics, temporal planning is tightly coupled with motion planning. The high-level temporal plan defines what to do and when, but each action (e.g., "move from A to B") requires a low-level trajectory that is geometrically and dynamically feasible. This creates a hierarchical planning problem where the temporal planner's abstract durations must be validated and refined by a motion planner that computes the actual paths and control commands.

In automated production lines, temporal planning schedules the concurrent execution of multiple robotic arms, grippers, and conveyors. It ensures tasks like welding, part placement, and quality inspection occur in the correct sequence with precise timing to maximize throughput and avoid collisions. For example, a planner must schedule a robot to pick up a component, move it to an assembly jig, and hold it in place for exactly 3.2 seconds while a second robot performs a riveting operation, all within a strict cycle time.

Self-driving cars use temporal planning to navigate complex traffic scenarios with dynamic actors. It reasons about time to collision, traffic light phases, and merging windows. The planner must generate a trajectory that not only avoids obstacles but also adheres to traffic laws defined by timing (e.g., a yellow light duration) and coordinates maneuvers like lane changes within safe temporal gaps between other vehicles. This extends to fleets of autonomous mobile robots in warehouses, where temporal planning coordinates their paths to prevent deadlock at intersections.

Mission planning for satellites and spacecraft is a quintessential temporal planning problem. Planners schedule:

• Communication windows with ground stations, which are constrained by orbital mechanics and visibility periods.
• Payload operations like sensor activation, which may have limited power budgets and thermal constraints.
• Maneuver execution for orbital adjustments, which must be timed precisely to conserve fuel. These plans often span days or weeks and must account for strict, immutable temporal constraints derived from celestial mechanics.

In robotic-assisted surgery, temporal planning can orchestrate the actions of multiple robotic tools and imaging systems. For a complex procedure, the planner might sequence:

• Tool insertion and retraction to minimize tissue trauma.
• Concurrent actuation of a cauterizing tool and a suction device.
• Synchronized movement with real-time imaging updates. The goal is to optimize the procedure's duration while respecting critical path dependencies and safety margins, reducing patient time under anesthesia.

Temporal planning optimizes the flow of goods in highly automated fulfillment centers and ports. It schedules:

• Loading/unloading of containers from ships, trucks, and trains, which have strict arrival and departure deadlines.
• Movement of autonomous guided vehicles (AGVs) across shared pathways, requiring precise timing to avoid congestion.
• Operation of cranes and sortation systems to meet shipping cut-off times. The planner must continuously re-optimize schedules in response to delays, maximizing resource utilization across a 24/7 operation.

Managing a modern electrical grid with volatile renewable sources (solar, wind) and storage requires fine-grained temporal planning. Algorithms schedule:

• Generator dispatch to meet predicted demand peaks.
• Battery charge/discharge cycles to arbitrage energy prices and stabilize the grid.
• Demand-response actions like temporarily adjusting industrial load, which must be executed at precise times to prevent blackouts. Plans are generated with horizons from minutes to days, balancing cost, reliability, and strict temporal constraints on equipment ramp-up times.

A hierarchical planning formalism that decomposes complex tasks into networks of subtasks using a library of pre-defined methods. Unlike classical planners that search for action sequences, HTN planners search through possible task decompositions until primitive, executable actions are reached. This approach is highly efficient for domains with known structure, such as manufacturing or logistics.

• Key Mechanism: Uses task reduction to replace abstract tasks with more concrete subtasks.
• Contrast with Temporal Planning: HTNs focus on logical decomposition; temporal planning adds explicit reasoning about action durations and concurrent execution.

The standardized, formal language used to model automated planning problems. It allows for the precise definition of predicates, actions, objects, and initial and goal states. Temporal extensions to PDDL, such as PDDL2.1, introduce constructs for modeling durative actions, continuous numeric change, and metric optimization (e.g., minimize total time).

• Core Components: Domain (predicates and actions) and Problem (objects, initial state, goal).
• Role in Temporal Planning: Serves as the input specification for temporal planners like OPTIC or TFD, defining the preconditions, effects, and duration of every action.

The process of computing a smooth, time-parameterized path that minimizes a cost function (e.g., energy, jerk, time) while satisfying dynamic constraints and avoiding collisions. It operates at a lower level than temporal planning, focusing on the continuous physics of movement.

• Key Input: A geometric path or a set of waypoints from a motion planner.
• Key Output: A time-optimal or energy-optimal trajectory specifying positions, velocities, and accelerations over time.
• Integration with Temporal Planning: A temporal plan's high-level actions (e.g., "Move from A to B in 5 seconds") are realized by a trajectory optimizer that computes the precise motor commands.

An advanced, online control method where a dynamic model of the system is used to predict future behavior over a finite horizon. It solves an optimization problem at each control step to determine the optimal immediate control inputs, then executes the first step and re-plans. This creates a feedback loop robust to disturbances.

• Core Principle: Receding horizon control with continuous re-optimization.
• Contrast with Temporal Planning: MPC handles low-level, continuous control and dynamics in real-time. Temporal planning provides the high-level action sequence and schedule that MPC's cost function can be tuned to achieve.

The critical processes that close the loop between planning and physical action. Execution monitoring observes the robot's state and environment to detect plan failures (e.g., an object is missing) or unexpected events. Replanning (or contingency planning) is then triggered to generate a new viable plan from the current state.

• Key Challenge: Distinguishing between transient sensor noise and a true plan failure.
• Temporal Aspect: In temporal domains, replanning must account for time already elapsed and new temporal constraints imposed by the changed situation.

A curated repository of reusable, parameterized behaviors that serve as the atomic building blocks for complex tasks. A skill (or motion primitive) is a short, robust sequence of control commands for a fundamental action like "pick," "place," or "move-to-pose."

• Abstraction Benefit: Allows high-level planners (temporal or HTN) to reason in terms of reliable skills rather than raw joint torques.
• Temporal Parameter: Each primitive has an associated duration and preconditions, which are the essential inputs for a temporal planner to sequence and schedule them.