Global Path Planner with Multi-Model Switching

arXiv:2609.13015 · cs.RO, cs.SE, cs.SY, eess.SY · Submitted 2026-09-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Global Path Planner with Multi-Model Switching".

Dev: Here's the gist of the paper. This work enhances global path planning via a pure-pursuit controller with multi-model kinematic switching that sustains plan fidelity across diverse terrains A core innovation is adaptive kinematic modeling, enabling real-time switching between kinematic models based on terrain features and robot states The system includes a traversability graph for terrain analysis, a Heading-Aware A∗ algorithm generating feasible paths, and a multi-model Pure Pursuit controller for dynamic tracking This adaptability optimizes path efficiency and energy use in challenging scenarios Path Planning Framework The path planning process begins with obtaining a raw point cloud of the environment from sensor measurements.

Rosa: This data is processed to compute traversability information through height filtering, ground segmentation using RANSAC, and surface normal estimation.

Dev: Navigable regions are further refined through coarser grid processing and validated with 2.5D maps.

Rosa: Obstacles are then grouped via spatial clustering using DBSCAN to mitigate noise and potential misclassification.

Paper summary: Rosa: So this paper, "Global Path Planner with Multi-Model Switching," it's about making global path planning better for robots by using a pure pursuit controller with multi-model kinematic switching to keep the plan accurate across different terrains.

Dev: It sounds like the main idea is this core innovation of adaptive kinematic modeling, which lets the system switch between different kinematic models in real time based on what the robot is doing and what kind of terrain it's on.

Taro: So, we're looking at a traversability graph to analyze terrain first, then using a Heading-Aware A* algorithm to find paths that are feasible for those conditions.

Rosa: Right, so the search algorithm isn't just finding a point; it’s searching in a state space that includes both the vertex position and the arrival heading.

Dev: That orientation dependency is key because it lets them add penalties for things like sharp turns and terrain slope right into the cost calculation during those A* searches.

Taro: And they have this Heading-Aware A* which calculates a cumulative cost based on step costs, including a node cost, an orientation change penalty, and a terrain interaction cost based on the slope angle theta.

Rosa: The orientation change penalty specifically is calculated as the weight of psi times the absolute difference between that edge's heading and your current heading.

Dev: And that terrain interaction cost scales with the slope angle theta, where larger angles mean a higher traversal cost for that segment.

Taro: So, they build this whole framework starting from a raw point cloud and processing it through height filtering, ground segmentation using RANSAC, and then spatial clustering to get a traversability graph.

Rosa: Once you have that graph, the actual path planning happens on this Heading-Aware A* search space to generate the optimal path.

Dev: The whole idea is that the system generates feasible paths for non-holonomic robots by incorporating orientation information into the search process, which helps create smooth trajectories even when things get tricky.

Taro: And then you take those paths and feed them into a multi-model Pure Pursuit controller for dynamic tracking, where the controller switches between unicycle and omnidirectional kinematic models.

Rosa: That switching is what keeps the fidelity high across diverse terrains, meaning it handles smooth transitions between different types of movement when it hits obstacles or changes elevation.

Dev: It’s interesting how they define these switching thresholds based on heading error and vertical displacement, which suggests a real-time decision-making process during tracking.

Taro: The paper tests this framework in simulation with things like the Artaban quadruped and an Xthree quadrotor drone to show that it works for different robot types.

Rosa: And they also have real-world experiments on the Artaban hardware where the robot successfully navigates a narrow doorway by switching its motion style.

Dev: So, what this means for us in terms of control loops, it's about creating a system that can adapt its fundamental motion model while maintaining tracking accuracy under changing conditions.

Taro: It shows how to make a path planner robust enough to handle elevation changes and open environments gracefully because it doesn't rely on just one fixed kinematic model.

Rosa: The overall result is a planning pipeline that combines terrain analysis, an A* search with heading awareness, and dynamic switching control for robust navigation.

Dev: I think the big implication here is moving beyond systems that assume a single motion type, which limits how well they can perform in varied environments like those we see outside the lab.

Taro: It gives us a blueprint for building autonomy where the robot can reason about and select the most suitable motion model for whatever task it's currently facing.

Rosa: So, this paper is really showing how to integrate perception, terrain analysis, and motion control into one adaptive system for better global path planning.

Dev: It’s a solid piece of work because it addresses the real-time challenge of having to adapt the motion model while keeping the tracking error down.

Taro: The authors acknowledge that this framework is designed to sustain plan fidelity across diverse terrains, but they also note that future work needs to focus on dynamic obstacle avoidance and online replanning in more challenging scenarios.

Rosa: It’s good to see them point out where it stops working right now, which tells you what the next steps for improving this system will be.

Dev: So, looking at the title, "Global Path Planner with Multi-Model Switching," it really captures that core concept of having a flexible planner that handles different motion types based on the environment.

Taro: And considering how they handle switching between models for things like narrow doorways and open spaces in their testing, it seems very practical for real-world applications right now.

Rosa: It does suggest that adding orientation information directly into the search cost function is a way to get smoother paths without needing overly complex, slow planning steps.

Dev: The whole structure feels like a strong control loop where the planning part informs the controller about what kind of motion model should be active next.

Taro: It’s definitely useful for anyone building systems that need to navigate environments with significant changes in ground texture or slope.

Conclusion: Rosa: So we're wrapping up on this paper, "Global Path Planner with Multi-Model Switching," which basically shows how to make a robot's long-distance path planning way more flexible by using a system that can instantly switch its movement style based on what it sees.

Dev: Yeah, the authors are focused on this adaptive kinematic modeling thing where the robot isn't stuck using just one way of moving, but changes between models in real time.

Taro: It's about having this whole graph search and a controller that can handle different terrains because they aren't assuming everything is the same everywhere.

Rosa: The core idea here is to get a path that works across totally different ground conditions, whether it's smooth pavement or really rough dirt.

Dev: And the way they do it, they build this heading-aware A star algorithm that already considers the robot's current heading when searching for the best route.

Taro: That orientation information in the search cost helps them generate paths that are actually smooth and feasible for a non-holonomic robot to follow.

Rosa: So, what this means for us is that we move away from rigid path planning where you just calculate one perfect line and hope it works everywhere.

Dev: It’s about making the path planner itself smarter, giving it the ability to react to the local environment and adjust its own internal model while moving.

Taro: That adaptability is crucial because in the real world, you don't know what kind of ground you're going to hit until you get there.

Rosa: And they tested this stuff on real hardware, showing that it actually works outside the lab for things like navigating tricky doorways and handling big elevation changes.

Dev: The results show it can maintain positional tracking consistently, which is a big deal when dealing with those kind of dynamic switching in motion control loops.

Taro: It proves that having a robust system means you don't just build one good planner; you build one that knows how to change its mind about its movement on the fly.

Rosa: It really highlights how important it is to integrate perception and planning so deeply into the control loop for real autonomy.

Pietro Gori, Francesco Iotti, Eduard Zelenay, *Rastislav Marko*, *Michele Pierallini*, *Franco Angelini*, *Gabriele Pannocchia*, *Manolo Garabini*

Centro di Ricerca “Enrico Piaggio” and the Dipartimento di Ingegneria dell’Informazione, University of Pisa · Panza Robotics · Institute of Robotics and Cybernetics, Faculty of Electrical Engineering and Information Technology, Slovak University of Technology in Bratislava · Dipartimento di Ingegneria Civile e Industriale, University of Pisa

cs.RO, cs.SE, cs.SY, eess.SY

Submitted: 2026-09-11

Updated: 2026-09-11

Project page: https://panzarobotics.com

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 83/100

The gist: This work enhances global path planning via a pure-pursuit controller with multi-model kinematic switching that sustains plan fidelity across diverse terrains A core innovation is adaptive kinematic…

Key concepts

Traversability Graph (TG)
This graph represents the navigable areas of the environment. Each node in the graph corresponds to a safe position for the robot, and edges connect adjacent nodes if movement between them is possible. The cost of each edge reflects factors like collision risk, distance, and slope, guiding the planner toward efficient routes.
Heading-Aware A* Algorithm
This search algorithm finds optimal paths by considering not just position but also the robot's arrival heading at each point. It incorporates an orientation change penalty into its cost calculation. This means the planner favors paths that require smoother turns or better alignment with the current direction, leading to more natural and feasible trajectories.
Multi-Model Pure Pursuit Controller
This controller tracks the planned path by switching between different motion models. It uses a unicycle model for planar motion and an omnidirectional model when needed. The switch happens based on how much the robot needs to change its heading or vertical position, allowing it to use the most appropriate movement strategy for the current situation.
Adaptive Kinematic Modeling
This is the core innovation where the system automatically selects between different ways a robot can move (kinematic models). It switches between a standard unicycle model and an omnidirectional model in real-time. This switching is triggered by analyzing local terrain features or the robot's current state, making the control system highly flexible.

Terminology

Summary

This work enhances global path planning via a pure-pursuit controller with multi-model kinematic switching that sustains plan fidelity across diverse terrains A core innovation is adaptive kinematic modeling, enabling real-time switching between kinematic models based on terrain features and robot states The system includes a traversability graph for terrain analysis, a Heading-Aware A∗ algorithm generating feasible paths, and a multi-model Pure Pursuit controller for dynamic tracking This adaptability optimizes path efficiency and energy use in challenging scenarios

Path Planning Framework

The path planning process begins with obtaining a raw point cloud of the environment from sensor measurements. This data is processed to compute traversability information through height filtering, ground segmentation using RANSAC, and surface normal estimation. p. 1

Where the paper says it

1 Traversability Map Initially, we obtain raw point cloud data of robot positions from sensor measurements. From these data, we identify collision-free areas based on terrain features such as height filtering, ground segmentation using RANSAC, and surface normal estimation.

Navigable regions are further refined through coarser grid processing and validated with 2. p. 1
Where the paper says it

Points are classified according to surface normals: vertical normals indicate navigable regions, horizontal normals represent obstacles, and intermediate cases are assessed using grid-based slope and height evaluations. Navigable regions are further refined through coarser grid processing and validated with 2.

5D maps. Obstacles are then grouped via spatial clustering using DBSCAN to mitigate noise and potential misclassification. p. 1
Where the paper says it

Navigable regions are further refined through coarser grid processing and validated with 2. Finally, obstacles are grouped via spatial clustering (DBSCAN [Deng, 2020]), and nearby navigable points are reclassified to mitigate noise and potential misclassification.

The traversability graph TG = (V, E) is built where each vertex encodes a feasible robot position with associated traversability cost c(vi). p. 1
Where the paper says it

Given the sets Cfree and Cobs, we build the traversability graph TG = (V, E), which serves as the discrete representation of the environment for path planning. Each vertex vi = (pi , c(vi), Ei) ∈ V encodes a feasible robot position, where pi ∈ Cfree denotes the spatial position, c(vi) is the traversability cost, and Ei ⊆ E is the set of incident edges of vertex i.

The cost function c(vi) is defined as the sum of collision, distance, and slope costs. p. 1
Where the paper says it

This graph structure enables efficient search for optimal paths, leveraging both geometric connectivity and traversability information. The cost function c : V → R for each vertex vi ∈ V is computed as follows: c(vi) = wcollccoll(vi) + wdistcdist(vi) + wslopecslope(vi), (1) where ccoll, cdist, cslope : V → R are the collision, distance, and slope cost functions, respectively, and wcoll, wdist, wslope are the corresponding weights.

Search Algorithm Details

The Heading-Aware A∗ Algorithm computes optimal paths by searching a state space that combines vertex position with arrival heading ψi. The cumulative cost g(vi, ψi) is determined by summing the step costs of individual transitions. p. 2

Where the paper says it

The cumulative cost, g(vi , ψi), to reach a state (vi , ψi) is determined by summing the costs of individual transitions. For each transition from a current state (vj , ψj ) to a successor state (vi , ψi), the step cost, Cstep, consists of three primary components: • Node Cost (Cbase): An inherent cost associated with traversing the edge ξ i j , representing the baseline effort required for movement.

The step cost Cstep consists of three primary components: Node Cost (Cbase), Orientation Change Penalty (Chead), and Terrain Interaction Cost (Cslope). p. 2
Where the paper says it

The cumulative cost, g(vi , ψi), to reach a state (vi , ψi) is determined by summing the costs of individual transitions. For each transition from a current state (vj , ψj ) to a successor state (vi , ψi), the step cost, Cstep, consists of three primary components: • Node Cost (Cbase): An inherent cost associated with traversing the edge ξ i j , representing the baseline effort required for movement.

The orientation change penalty is calculated as Chead = wψ · ∆ψ where ∆ψ is the absolute difference between the edge orientation and the current heading. p. 2
Where the paper says it

• Orientation Change Penalty (Chead): A penalty for deviations in the robot’s heading. The orientation of the traversed edge, ψedge, is computed from the displacement vector between pj and pi .

The terrain interaction cost Cslope increases the cost for traversing steeper terrain based on the slope angle θ and direction relative to the surface normal. p. 2
Where the paper says it

(4) • Terrain Interaction Cost (Cslope): A terrainadaptive cost that penalizes movements based on the local slope and the travelling direction relative to the surface normal. This cost is calculated as: Cslope = wslope · sin(θ) · cos(ϕ), (5) where wslope is a weighting factor, θ is the slope angle at vj (3), and ϕ is the angle between the normalized movement vector (from pj to pi) and the surface normal at vj , i.

Trajectory Tracking Control

The Multi-Model Pure Pursuit Controller tracks the path generated by Heading-Aware A∗ and computes velocity commands. The controller dynamically switches between unicycle and omnidirectional kinematic models based on predefined thresholds for heading error ∆ψ and vertical displacement zdisp. p. 3

Where the paper says it

(14) From eR, the planar distance d, heading error ∆ψ, and vertical displacement zdisp are computed as d = q e 2 x + e 2 y , (15) ∆ψ = atan2(ey, ex), (16) zdisp = ez. (17) The controller selects the unicycle model when |∆ψ| < δψ and |zdisp| < δz, (18) otherwise, the omnidirectional model is used.

For planar motion, the unicycle model uses linear velocity proportional to the distance error d and angular velocity proportional to the heading error ∆ψ. p. 3
Where the paper says it

(20) Unicycle Control In unicycle mode, the linear velocity is proportional to the distance error: u uni = kvd cos(∆ψ), (21) saturated to the admissible velocity range [−u uni max, uuni max]. The angular velocity is computed from the heading error: ω = kω∆ψ, (22) where kv and kω are proportional gains for linear and angular velocity, respectively.

In omnidirectional mode, the commanded velocity u omni is calculated based on minimizing tracking error D. p. 3
Where the paper says it

After selecting the appropriate kinematic model, the controller computes the velocities required to reach the lookahead point. Omnidirectional Control In omnidirectional mode, the commanded velocity is proportional to the tracking error: u omni = 0, D < ϵ, S eR, otherwise, (19) where S = min u omni max D , 1 , D = ∥eR∥2.

Validation and Results

The proposed path planner was validated in simulation using the Artaban quadruped and an X3 quadrotor drone. In simulation, the robot successfully navigated a narrow doorway by adopting unicycle-like motion and then transitioned to omnidirectional movements in open space. p. 4

Where the paper says it

During the door traversal, the robot adopts a unicycle-like motion, aligning its longitudinal axis with the doorway to ensure successful passage. Once through the door and into the open space, the robot transitions to more flexible, omnidirectional movements, leveraging its full kinematic capabilities to efficiently reach the goal.

Real-world experiments on the Artaban hardware demonstrated that the robot successfully completed the path with positional tracking consistent with simulation results. p. 4
Where the paper says it

2 Real-World Experiment To further evaluate the robustness of our approach, we conducted experiments on the real Artaban hardware. The results demonstrate that the robot successfully completed the path, validating the effectiveness of the planner in a real-world setting.

The framework was shown to be robust in managing elevation changes and open environments. p. 5
Where the paper says it

The planner generates feasible and smooth paths by incorporating orientation information into the graph search process, while the controller dynamically switches between unicycle and omnidirectional motion models according to the local terrain and path geometry. The framework was validated in simulation and real-world experiments using the Artaban robot, demonstrating reliable navigation through narrow passages, elevation changes, and open environments.

Conclusion

The integrated framework combines a Heading-Aware A∗ planner with a multi-model pure pursuit controller for adaptive trajectory tracking. This approach generates feasible and smooth paths by incorporating orientation information into the graph search process and dynamically switching between motion models according to local terrain and path geometry. p. 5

Where the paper says it

reference multi-model pure pursuit controller for adaptive trajectory tracking. The planner generates feasible and smooth paths by incorporating orientation information into the graph search process, while the controller dynamically switches between unicycle and omnidirectional motion models according to the local terrain and path geometry.

The system was validated in simulation and real-world experiments using the Artaban robot, demonstrating reliable navigation through narrow passages, elevation changes, and open environments. p. 5
Where the paper says it

The planner generates feasible and smooth paths by incorporating orientation information into the graph search process, while the controller dynamically switches between unicycle and omnidirectional motion models according to the local terrain and path geometry. The framework was validated in simulation and real-world experiments using the Artaban robot, demonstrating reliable navigation through narrow passages, elevation changes, and open environments.

Future work will focus on dynamic obstacle avoidance and online replanning in more challenging scenarios. p. 5
Where the paper says it

The framework was validated in simulation and real-world experiments using the Artaban robot, demonstrating reliable navigation through narrow passages, elevation changes, and open environments. Future work will focus on dynamic obstacle avoidance and online replanning in more challenging scenarios.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive kinematic modeling allows for real-time switching between kinematic models based on terrain features and robot states, which optimizes path efficiency and energy use in challenging scenarios. The system enables real-time switching between kinematic models based on terrain features and robot states.

  2. A novel Heading-Aware Astar algorithm incorporates orientation costs into path planning, enabling smoother and more efficient trajectories. This enhancement ensures that the generated paths are not only geometrically feasible but also align with the robot's desired orientation constraints throughout the navigation process.

  3. The multi-model pure pursuit controller dynamically switches between unicycle and omnidirectional kinematic models according to whether planar or nonplanar motion is required, ensuring the robot always follows the most efficient control strategy. This adaptability allows for effective traversal through narrow doorways by adopting a unicycle-like motion and then transitioning to more flexible movements.

  4. The path planning pipeline integrates raw point cloud data processing with traversability graph construction using spatial clustering (DBSCAN) and cost functions that penalize collision, distance, and slope. This robust preprocessing step ensures the resulting graph accurately represents all possible positions in the environment by classifying navigable regions and obstacles effectively.

  5. The heading-aware Astar algorithm includes a penalty for orientation change, calculated as Chead = wpsi · ∆ψ, which generates smooth trajectories by penalizing sharp turns and terrain slope effects during path search. This feature directly contributes to generating smooth and feasible trajectories also for non-holonomic robots.

Abstract

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.

Related papers