Source-linked AI summary
Global Path Planner with Multi-Model Switching
Pietro Gori, Francesco Iotti, Eduard Zelenay, Rastislav Marko, Michele Pierallini, Franco Angelini, Gabriele Pannocchia, Manolo Garabini
TL;DR
Robots often depend on human operators, limiting autonomy when intervention is impractical or hazardous. This work combines terrain-aware path planning with a Pure Pursuit controller that switches kinematic models. It reports reliable tracking and navigation across simulation and real-world Artaban experiments, with an X3 drone simulation also demonstrating accurate trajectory following.
Problem
Robots’ dependence on human operators limits their autonomy in scenarios where intervention is impractical or hazardous.
Source: §1 · p. 1Method
The framework combines a traversability graph and Heading-Aware A* planner with a Pure Pursuit controller that switches kinematic models according to terrain and path geometry.
Source: §2.1 · p. 2, §2.3 · p. 4, §4 · p. 6Results
The planner generated accurately followed trajectories in X3 simulation and demonstrated reliable Artaban navigation in simulation and real-world experiments.
Source: X3 quadrotor drone · p. 5, §4 · p. 6Takeaways & Limitations
The framework supports reliable navigation through narrow passages, elevation changes, and open environments using adaptive tracking.
Source: §4 · p. 6
Abstract
from arXiv · showhide
This work enhances global path planning via a pure-pursuit controller with multi-model kinematic switching that sustains plan fidelity across diverse terrains. The system includes a traversability graph for terrain analysis, a Heading-Aware A* algorithm for generating feasible paths, and a multi-model Pure Pursuit controller for dynamic tracking. A core innovation is adaptive kinematic modeling, enabling real-time switching between kinematic models based on terrain features and robot states. This adaptability optimizes path efficiency and energy use in challenging scenarios. We validate the approach in simulation on different platforms, namely the Artaban quadruped and the X3 quadrotor drone, showcasing improved performance, robustness, and adaptability over standard baselines.
1. INTRODUCTION
The paper targets the robustness limits of fixed-kinematic-model planning with a framework that switches models to suit hardware, task, and environment. It combines a traversability graph, Heading-Aware A*, and Pure Pursuit tracking.
1. INTRODUCTION
- Fixed kinematic models can limit robustness across diverse environments, motivating a planner that switches models according to hardware constraints, task requirements, and environmental conditions.The contributions also include orientation-aware path costs and an adaptive trajectory follower.Source: §1 · p. 2
2. PATH PLANNING ALGORITHM
- The planning pipeline processes a point cloud into traversability information, searches for a path with Heading-Aware A*, and uses Pure Pursuit to refine it into a smooth trajectory.Figure 1 illustrates the pipeline.Source: §2 · p. 2
2.1 Traversability Map
- Terrain processing classifies traversable and obstacle regions, then represents feasible positions and collision-free transitions as a graph for path search.Graph vertices encode position and traversability cost; the cost combines collision, distance, and slope terms.Source: §2.1 · p. 2
Collision Cost
Collision cost penalizes proximity to obstacles while retaining clearance in narrow passages.
Collision Cost
- The collision cost penalizes traversable vertices according to their proximity to obstacle regions.A robot-sized disk around the vertex is used to test for hard collision.Source: Collision Cost · p. 2
- An obstacle inside the robot’s XY circumscribed disk is treated as a hard collision.The disk radius is based on the robot’s length and width.Source: Collision Cost · p. 2
- Otherwise, a distance-dependent soft penalty encourages clearance while remaining permissive in narrow passages.The penalty uses the minimum Euclidean distance to Cobs; A sets its maximum and σ its decay rate.Source: Collision Cost · p. 2
Distance Cost
The distance cost measures the Euclidean distance between a graph vertex and the next vertex.
Distance Cost
- The distance cost cdist(vi) is the Euclidean distance from vertex position pi to the next vertex position pj.It measures the length of the next graph transition.Source: Distance Cost · p. 2
- The cost is defined between consecutive vertex positions in the traversability graph.The passage identifies pi as the current vertex position and pj as the next one.Source: Distance Cost · p. 2
- cdist(vi) expresses the spatial separation from vi to its next graph vertex, rather than a terrain-slope or obstacle-clearance term.The passage defines it as Euclidean distance between the two positions.Source: Distance Cost · p. 2
Slope Cost
The planner assigns greater traversal cost to steeper terrain, incorporates heading and terrain penalties into A* search, and switches Pure Pursuit kinematic models according to path geometry and robot state.
Slope Cost
- Steeper terrain receives higher slope cost, with local slope estimated from a least-squares plane fitted to neighboring points.The slope angle is measured relative to the vertical axis.Source: Slope Cost · pp. 2–3
2.2 Heading-Aware A* Algorithm
- Heading-Aware A* searches states that pair graph vertices with arrival headings, penalizing orientation changes and terrain interactions while allowing steep slopes when needed.Its transition cost combines baseline node cost, heading-change penalty, and a slope term based on slope angle and movement direction relative to the surface normal.Source: §2.2 · p. 3
2.3 Multi-Model Pure Pursuit Controller
- Pure Pursuit selects a lookahead point and switches between unicycle and omnidirectional models according to path geometry and robot state.The unicycle model is selected when heading error and vertical displacement are below predefined thresholds; otherwise, the omnidirectional model is used.Source: §2.3 · p. 4
Omnidirectional Control
In omnidirectional mode, commanded velocity is proportional to tracking error.
Omnidirectional Control
- In omnidirectional mode, commanded velocity is proportional to tracking error.Source: Omnidirectional Control · p. 4
Unicycle Control
In unicycle mode, linear velocity follows distance error and angular velocity follows heading error.
Unicycle Control
- In unicycle mode, linear velocity is proportional to distance error, while angular velocity is computed from heading error.Source: Unicycle Control · p. 4
3. NUMERICAL RESULTS
The path planner is evaluated through simulation and real-world experiments, with the Artaban robot and an X3 quadrotor used for validation.
3. NUMERICAL RESULTS
- The planner is evaluated in simulation and real-world experiments to assess feasible, efficient path generation across environments.Source: §3 · p. 4
3.1 Simulation
- The validation platforms are the Artaban robot and an X3 quadrotor drone.Source: §3.1 · p. 4
Artaban quadruped
In simulation, Artaban traverses a narrow doorway before entering open space, switching from unicycle-like to omnidirectional motion. Position and orientation references closely match ground truth, with greater differences in vertical motion and body tilt.
Artaban quadruped
- The Artaban is a hazardous-environment legged robot with cardan joints, four-bar linkages, and three actuators per leg.Its driver uses PD controllers and inverse dynamics for rear-leg kinematic loops, with friction and inertia compensation.Source: Artaban quadruped · p. 4
- The global planner node interfaces with Quad-SDK and defines the robot’s position and velocity for trajectory following.Source: Artaban quadruped · p. 5
- Position and orientation references closely match ground truth, especially in X, Y, and yaw; Z oscillations and body tilts differ from the planner’s smoother trajectories.The passage attributes the Z oscillations to physical dynamics and terrain interaction, and the larger tilts to the simulated robot’s natural motion.Source: Artaban quadruped · p. 5
- Artaban aligns its longitudinal axis with the narrow doorway using unicycle-like motion, then switches to omnidirectional movement in open space.This scenario demonstrates adapting motion strategy to environmental constraints.Source: Artaban quadruped · p. 5
X3 quadrotor drone
In simulation, the X3 drone navigates a complex environment and is reported to follow computed trajectories accurately. The paper also reports successful real-world path completion and tracking on Artaban, with minor discrepancies attributed to sensor noise and execution imperfections.
X3 quadrotor drone
- The X3 drone successfully follows trajectories through a complex environment requiring position and orientation control to avoid obstacles and reach the target.Source: X3 quadrotor drone · p. 5
3.2 Real-World Experiment
- Real Artaban hardware completed the path, with positional tracking consistent with simulation but minor discrepancies from sensor noise and execution imperfections.Orientation tracking is inferred to be similar to simulation, with real-world dynamics slightly amplifying discrepancies.Source: §3.2 · p. 5
4. CONCLUSION
The conclusion presents an integrated planner and adaptive trajectory follower for complex planar and non-planar navigation. Validation on Artaban covers simulation and real-world environments, while dynamic obstacle avoidance and online replanning remain future work.
4. CONCLUSION
- The framework combines a traversability-graph Heading-Aware A* planner with a multi-model Pure Pursuit controller for adaptive trajectory tracking.The planner incorporates orientation into graph search; the controller switches between unicycle and omnidirectional models according to local terrain and path geometry.Source: §4 · pp. 5–6
- Artaban validation in simulation and real-world experiments demonstrated reliable navigation through narrow passages, elevation changes, and open environments.Source: §4 · p. 6
- Dynamic obstacle avoidance and online replanning in more challenging scenarios are identified as future work.Source: §4 · p. 6