# Smac Planner Hybrid A Star > --- name: smac-planner-hybrid-a-star description: 'Configure Nav2 SMAC Hybrid-A* planner. Use when planning kinematically feasible paths with minimum turning radius constraints.' --- # SMAC Hybrid-A* Planner ## Plugin ``` nav2_smac_planner::SmacPlannerHybridAstar ``` ## Algorithm Hybrid-A* extends A* into the SE2 state space (x, y, theta), producing paths that respect non-holonomic motion constraints. The search graph discretizes heading into angular bins and expands nodes using Dubins or Reeds-Shepp motion primitives. An analytic expansion periodically attempts to connect the current node directly to the goal using a closed-form curve, dramatically reducing search time when a clear path exists. The planner searches in a 3D space: (x_cell, y_cell, heading_bin). Each expansion applies a kinematically feasible motion primitive, ensuring the output path can be physically executed by the robot. ## Critical: Minimum Turning Radius ```yaml minimum_turning_radius: 0.20 # meters - for differential drive robots ``` **For differential drive robots**, the true minimum turning radius is effectively zero (the robot can spin in place). However, setting it to exactly 0 produces degenerate motion primitives. Use a small value like **0.1–0.4 meters** to produce smooth, natural-looking arcs while still allowing tight turns. Larger values (0.5+) force wider arcs and may cause failures in tight spaces. For Ackermann (car-like) robots, set this to the actual mechanical minimum turning radius. ## Motion Models | Model | Description | Use Case | |-------|-------------|----------| | `DUBIN` | Forward-only arcs (left, straight, right) | Robots that cannot reverse (e.g., car-like robots without reverse gear) | | `REEDS_SHEPP` | Forward and reverse arcs | Robots that can reverse. Produces more flexible paths, especially in constrained spaces like parking/docking. | For differential drive robots, **Reeds-Shepp** is preferred since the robot can move backward. For Ackermann robots, choose based on whether reversing is allowed in your application. ## Key Parameters | Parameter | Type | Default | Description | |-----------|------|---------|-------------| | `minimum_turning_radius` | double | 0.40 | Minimum turning radius in meters. | | `max_iterations` | int | 1000000 | Maximum SE2 node expansions. | | `max_planning_time` | double | 5.0 | Hard time limit in seconds. | | `motion_model_for_search` | string | "DUBIN" | "DUBIN" or "REEDS_SHEPP". | | `angle_quantization_bins` | int | 72 | Number of heading bins. 72 = 5° resolution, 36 = 10°. More bins = finer heading control but larger search space. | | `lookup_table_size` | double | 20.0 | Size of the precomputed heuristic lookup table in meters. Should be ≥ the expected map size. | | `allow_unknown` | bool | true | Plan through unknown space. | | `allow_reverse_expansion` | bool | false | Only used with REEDS_SHEPP. Allow the planner to expand reverse motions. | ## Analytic Expansion Parameters The analytic expansion attempts to shortcut the search by computing a closed-form Dubins/Reeds-Shepp curve from the current node to the goal. | Parameter | Type | Default | Description | |-----------|------|---------|-------------| | `analytic_expansion_ratio` | double | 3.5 | How often (relative to search progress) to attempt analytic expansion. Lower = more frequent attempts. | | `analytic_expansion_max_length` | double | 3.0 | Maximum length of the analytic expansion in meters. | ## Penalty Parameters These penalties shape path characteristics by adding cost to certain types of motion. | Parameter | Type | Default | Effect | |-----------|------|---------|--------| | `change_penalty` | double | 0.0 | Cost for switching between forward and reverse motion. Higher values reduce oscillation. | | `non_straight_penalty` | double | 1.20 | Cost multiplier for turning vs going straight. Higher = straighter paths. | | `cost_penalty` | double | 2.0 | Multiplier for costmap cell costs. **Higher values make the planner strongly avoid high-cost areas** (near obstacles). Critical for safety margin tuning. | | `reverse_penalty` | double | 2.0 | Cost multiplier for reverse motion. Higher = prefers forward motion. | **Penalty tuning strategy:** - Robot hugs walls → increase `cost_penalty` (try 3.0–5.0). - Path has too many unnecessary turns → increase `non_straight_penalty`. - Robot reverses when it shouldn't → increase `reverse_penalty` or disable `allow_reverse_expansion`. - Robot oscillates forward/reverse → increase `change_penalty`. ## Complete YAML Configuration ```yaml planner_server: ros__parameters: expected_planner_frequency: 20.0 planner_plugins: ["GridBased"] GridBased: plugin: "nav2_smac_planner::SmacPlannerHybridAstar" tolerance: 0.25 max_iterations: 1000000 max_on_approach_iterations: 1000 max_planning_time: 5.0 minimum_turning_radius: 0.20 motion_model_for_search: "REEDS_SHEPP" angle_quantization_bins: 72 analytic_expansion_ratio: 3.5 analytic_expansion_max_length: 3.0 lookup_table_size: 20.0 allow_unknown: true allow_reverse_expansion: false cost_travel_multiplier: 2.0 change_penalty: 0.0 non_straight_penalty: 1.20 cost_penalty: 2.0 reverse_penalty: 2.0 downsample_costmap: false downsampling_factor: 1 cost_penalty: 2.0 smoother: max_iterations: 1000 w_smooth: 0.3 w_data: 0.2 tolerance: 1.0e-10 ``` ## When To Use **Good fit:** - You want kinematically feasible paths directly from the planner, reducing reliance on the controller to fix infeasible segments. - Ackermann steering robots with a real minimum turning radius. - Differential drive robots in constrained environments where smooth, arc-based paths reduce controller effort. - Docking or parking maneuvers where approach angle matters. **Poor fit:** - Simple open environments where NavFn + controller smoothing suffices (Hybrid-A* is slower). - Real-time replanning at high frequency (the 3D search is inherently more expensive). - Holonomic robots that don't benefit from non-holonomic motion planning. ## Performance Considerations Hybrid-A* is significantly slower than 2D planners because the search space is O(X × Y × θ) instead of O(X × Y). Mitigation strategies: - **Downsample the costmap** (`downsample_costmap: true`, `downsampling_factor: 2`). - **Reduce `angle_quantization_bins`** to 36 (10° resolution) if fine heading control isn't needed. - **Lower `max_planning_time`** and accept occasional planning failures that trigger replanning. - **Reduce `lookup_table_size`** if your map is smaller. ## Troubleshooting - **Planner times out in open space:** The lookup table or analytic expansion may be misconfigured. Increase `analytic_expansion_ratio` to attempt shortcuts more often. - **Paths have unnecessary reversal segments:** Set `allow_reverse_expansion: false` or increase `reverse_penalty`. - **Planner fails in narrow corridors:** Decrease `minimum_turning_radius`. Check that costmap inflation isn't closing off the corridor. - **Path doesn't match robot's actual capability:** Ensure `minimum_turning_radius` matches your robot. For diff-drive, 0.1–0.4m is typical.