Replanning

Replanning is triggered by environmental changes or execution failures that invalidate the current plan's assumptions. This requires continuous execution monitoring to detect events such as:

• Unexpected obstacles appearing in a planned path.
• Resource unavailability (e.g., a tool is missing).
• Dynamic goal updates from a human operator or higher-level system.
• Action failure where a commanded movement does not achieve its intended effect. The system must then reassess the current world state and compute a new plan.

Effective replanning is not a full restart. It leverages the current state and often reuses parts of the previous plan or search effort to find a new solution quickly. Key techniques include:

• Incremental planners that modify the existing plan graph.
• Anytime algorithms that can return a feasible solution quickly and improve it if time allows.
• Replanning heuristics that bias the search based on the previous successful path. This efficiency is crucial for real-time systems where computational latency directly impacts operational safety and performance.

Replanning often occurs at multiple levels of abstraction. A failure in a low-level motion plan (e.g., a collision) may be resolved locally, while a failure in a high-level task plan (e.g., 'assemble part') may require re-decomposition.

• High-Level (Task): A Hierarchical Task Network (HTN) planner might select a different method to achieve the same abstract goal.
• Mid-Level (Skill): The system might choose an alternative motion primitive from a skill library.
• Low-Level (Control): A Model Predictive Control (MPC) controller continuously re-optimizes the trajectory over a short horizon. This separation of concerns prevents minor execution issues from causing complete task replanning.

Replanning relies on an accurate and frequently updated world model or state representation. This model fuses sensor data to maintain a belief about:

• Object positions and properties (3D scene understanding).
• The robot's own configuration and dynamics.
• The status of other agents (multi-agent state). When a discrepancy is detected between the model's prediction and sensor observation, it triggers the replanning process. The fidelity of this model determines the replanner's ability to generate feasible new plans.

A new plan must satisfy all original hard constraints and safety protocols. The replanner must ensure:

• Collision avoidance with both static and dynamic obstacles.
• Adherence to kinematic and dynamic limits of the robot.
• Respect for temporal constraints and deadlines.
• Maintenance of stability criteria (e.g., for legged robots). Techniques like constraint-aware sampling in RRT* or feasibility checks within trajectory optimization are employed to generate only valid plans. Plan validation via simulation may be used before execution.

While reactive, replanning remains directed toward the original or updated high-level goal. However, it can also be opportunistic—seizing chances to improve efficiency. For example:

• If an obstacle clears, replanning might compute a more direct path.
• If a resource becomes available earlier than expected, the plan can be re-sequenced. This balances reactivity with goal commitment, ensuring the system is not perpetually re-optimizing but can capitalize on beneficial changes in the environment.

The foundational process of breaking down a complex, high-level goal into a structured sequence or hierarchy of simpler, actionable subtasks. This is the first step before motion planning can occur.

• Hierarchical Structure: Creates a tree of tasks, where high-level objectives are recursively split.
• Example: The goal 'Assemble a table' decomposes into 'Fetch leg', 'Align leg with tabletop', 'Insert fastener', and 'Tighten fastener'.
• Relation to Replanning: If a subtask fails (e.g., 'Fetch leg' because the leg is missing), replanning must occur at the appropriate level of the decomposition.

The algorithmic core that computes a sequence of valid configurations for a robot to move from a start to a goal while avoiding obstacles and respecting physical constraints (kinematics, dynamics).

• Core Problem: Find a path through the robot's configuration space (C-Space) from point A to point B.
• Key Algorithms: Includes sampling-based planners like RRT (Rapidly-exploring Random Tree) and PRM (Probabilistic Roadmap), and optimization-based methods.
• Relation to Replanning: Replanning is triggered when the initially computed motion plan becomes invalid due to a moving obstacle or a changed goal.

The real-time process of observing a robot's state and its environment during plan execution to detect deviations from expected outcomes, failures, or unexpected events.

• Sensors & State Estimation: Uses lidar, cameras, and encoders to compare the expected state from the plan with the actual, measured state.
• Failure Detection: Identifies conditions like actuator failure, object slippage, or the sudden appearance of a human in the workspace.
• Relation to Replanning: This is the trigger mechanism for replanning. Monitoring provides the 'why' and the new world state that necessitates a new plan.

A powerful planning formalism that decomposes high-level tasks into subtasks using a library of pre-defined 'methods' until primitive, executable actions are reached.

• Method Libraries: Encode domain expertise (e.g., one method to 'Navigate to Room' might involve 'Open Door' then 'Move Through').
• vs. Classical Planning: More expressive for complex, real-world tasks than pure STRIPS/PDDL representations.
• Relation to Replanning: An HTN planner can be used for the replanning step itself, taking the current (failed) state as a new initial state and re-invoking its decomposition methods.

An advanced, online control strategy that uses a dynamic model of the system to predict future behavior over a short horizon, solves an optimization problem for the best control sequence, and executes only the first step before repeating.

• Receding Horizon: Continuously re-plans at a high frequency (e.g., 100 Hz).
• Handles Constraints: Explicitly accounts for actuator limits, obstacle boundaries, and dynamics.
• Relation to Replanning: MPC is a form of continuous, low-level replanning. It inherently adapts to disturbances and changing conditions within its prediction window.

The process of computing a smooth, time-parameterized path (a trajectory) that minimizes a cost function (e.g., energy, time, jerk) while satisfying dynamic constraints and avoiding collisions.

• Beyond Geometry: Specifies not just the path, but also velocities, accelerations, and torques.
• Methods: Includes direct collocation, shooting methods, and optimization libraries like Crocoddyl or TrajOpt.
• Relation to Replanning: After a high-level replanning step selects a new sequence of actions, trajectory optimization is often used to generate the precise, executable motion for the new plan.