Replanning

Recomputing a path or set of steps mid-execution in response to new information, instead of sticking to the original plan.

Replanning means a robot recomputes its remaining path, trajectory, or task steps during execution because the environment changed, execution failed, or new information arrived. A plan computed in advance assumes a static world, but in reality new obstacles appear, objects get bumped out of place, and grasps fail — continuing with the original plan in these cases leads to failure. It happens at every level: in navigation, the ROS 2 navigation stack Nav2's default behavior tree recomputes the global path periodically at 1 Hz; in control, model predictive control re-solves a short future window every cycle, which can be seen as very high-frequency replanning; at the task level, Inner Monologue feeds success detection and scene descriptions back to a large language model so it can adjust its next step after a failure.

ExampleA mobile robot navigating to the kitchen with Nav2 is on its way when someone moves a chair into the hallway; once the costmap updates, the next replanning cycle produces a new route around the chair, and if no path can be found it triggers recovery behaviors such as clearing the costmap, rotating in place, or backing up.

Also called

Online Replanning, Dynamic Replanning

Related

Failure Recovery · Model Predictive Control · D* / D* Lite · Closed-loop Control · Behavior Tree · LLM-based Task Planning

Sources

Nav2 默认行为树 navigate_to_pose_w_replanning_and_recovery.xml（1 Hz 重规划） (Chinese)
Inner Monologue: Embodied Reasoning through Planning with Language Models