nav2_planners

Use SmacPlannerHybrid or SmacPlannerLattice for Ackermann and car-like robots because they generate kinematically feasible paths. Hybrid-A* supports Dubins (forward only) and Reeds-Shepp (forward and reverse) motion models with a configurable minimum turning radius.

NavFn uses Dijkstra or A* to build a navigation potential field and extracts paths by gradient descent, while SmacPlanner2D is a templated grid A* with path smoothing and multi-resolution support. Both suit circular robots, but Smac 2D avoids NavFn's path discontinuity artifacts.

Increase the w_traversal_cost parameter so the planner penalizes cells near obstacles more strongly, pushing paths toward open space at the cost of more expansions. Pair it with gentle costmap inflation, such as a cost_scaling_factor around 10.0.

Yes, set motion_model_for_search to REEDS_SHEPP to allow forward and reverse motion segments. You can then tune reverse_penalty, change_penalty, and retrospective_penalty to control how often the planner chooses backing maneuvers.

Use the Route Server for structured environments like warehouses or hospitals where navigation follows pre-defined graphs with semantic rules such as speed zones or one-way lanes. It runs Dijkstra on a GeoJSON graph in well under a millisecond even for large graphs.

Enable cache_obstacle_heuristic to reuse the obstacle heuristic between replans, which gives roughly a 40x speedup. You can also enable downsample_costmap for multi-resolution search and raise max_planning_time limits if expansions are being cut short.