Integrating Planning and Predictive Control Using the Path Feasibility Governor

summary

The gist

The Path Feasibility Governor proposes a framework for integrating path planners with nonlinear Model Predictive Control to generate dynamically feasible…

In short

The Path Feasibility Governor (PathFG) integrates path planning with Model Predictive Control (MPC) to create dynamically feasible, collision-free trajectories in complex, non-convex environments. By manipulating the reference passed to the MPC, PathFG ensures stability and constraint satisfaction while allowing path planners to handle difficult obstacle avoidance.

Key concepts

Path Planning vs. Path Tracking
Motion planning is split into two parts: finding a safe route (path planning) and following that route while respecting vehicle dynamics (path tracking). The problem is that traditional methods separate these, leading to a disconnect where the planner ignores the system's actual movement constraints.
Path Feasibility Governor (PathFG)
PathFG is a modular system that guides the MPC controller using intermediate references generated by a path planner. This integration ensures that the MPC always operates on paths that are guaranteed to be collision-free and physically achievable, maintaining stability even in non-convex settings.
Region of Attraction (ROA)
The ROA defines all possible starting states from which the closed-loop system can successfully reach a target reference. PathFG significantly expands this region, allowing the system to operate safely from any state that has a feasible path to the desired goal, not just those reachable within a short prediction window.
Input-to-State Stability (ISS)
ISS is a stability framework used here to prove that the combined PathFG and MPC system is stable. It mathematically guarantees that if the system receives an external input (like a reference), its state will remain bounded and converge towards a stable equilibrium point.

Terminology used across episodes

This episode discusses

The paper

Integrating Planning and Predictive Control Using the Path Feasibility Governor · Read on arXiv

Shu Zhang, James Y. Z. Liu, Dominic Liao-McPherson

University of British Columbia

Transcript

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

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Integrating Planning and Predictive Control Using the Path Feasibility Governor".

Dev: Here's the gist of the paper. The Path Feasibility Governor proposes a framework for integrating path planners with nonlinear Model Predictive Control to generate dynamically feasible, collision-free trajectories in non-convex environments Problem Decomposition and Motivation Motion planning is often decomposed into path planning and path tracking to improve tractability, but this decomposition introduces a kinodynamic disconnect where path planners may ignore system dynamics The PathFG addresses this by manipulating the reference passed to the MPC controller, guiding it along a path while ensuring constraint satisfaction, stability, and recursive feasibility This integration is challenging because traditional methods either lack theoretical guarantees or sacrifice modularity when incorporating controllers into the planning process Path Feasibility Governor Architecture The PathFG is a modular system that enables the integration of predictive controllers with a wide array of path planners Given a reference r, the path planner generates a collisionfree path, and the PathFG creates intermediate references that guide the system while ensuring they are feasible for the MPC This architecture offers several key advantages, including being computationally cheap and ensuring constraint satisfaction, asymptotic stability, and recursive feasibility of the MPC It also substantially enlarges the region of attraction of the closed-loop system to encompass all initial states from which a feasible path can be generated Theoretical Guarantees The theoretical analysis proves that the combined PathFG+MPC feedback policy satisfies both safety and asymptotic stability control objectives Theorem 1 establishes recursive feasibility and safety for any target reference r in Rε and initial state x0 in Dx(r) Theorem 2 proves asymptotic stability, showing that the point (¯xr, 1) is an asymptotically stable equilibrium point for the closed-loop system The proof relies on Input-to-State Stability (ISS) of the combined system and a finite-time convergence argument for the auxiliary reference s to 1 Numerical Validation and Performance Numerical validation using quadrotor navigation simulations demonstrates that the PathFG framework guarantees safety, achieves finite-time convergence to the target reference, and significantly reduces the computational burden of the MPC The results show that while ungoverned MPC requires a minimum prediction horizon of 20 to maintain feasibility, the combined PathFG+MPC controller successfully navigates obstacle-dense environments with a short prediction horizon N = 5 The PathFG component alone maintains an average computation time under 0.

Title and authors: Rosa: We've covered the title and authors of this paper, "Integrating Planning and Predictive Control Using the Path Feasibility Governor," and what it means in terms of tackling the motion planning problem. We talked about how they propose using a framework called PathFG to connect path planners with nonlinear Model Predictive Control one <ref:2507.09134#pg1,path planners with nonlinear Model Predictive Control>.

Dev: And we focused on how that framework manipulates the reference passed to the MPC controller, guiding it along a path while ensuring constraint satisfaction, stability, and recursive feasibility one <ref:2507.09134#pg1,manipulates the reference passed to the MPC controller, guiding it along a>.

Rosa: It’s interesting because they aren't just doing a simple handover between planning and tracking; they are actively guiding the controller with intermediate references one <ref:2507.09134#pg1>.

Dev: And that intermediate reference selection is what keeps the optimal control problem under the hood from becoming infeasible over time, which is key to maintaining recursive feasibility one <ref:2507.09134#pg1>.

Rosa: So, it’s about ensuring that even when you're following a pre-calculated path, the system dynamics still allow for a solution at every single step of that path one <ref:2507.09134#pg1>.

Dev: That’s right. It addresses the kinodynamic disconnect where planners ignore system dynamics by embedding the planning constraints directly into how the controller operates one <ref:2507.09134#pg1>.

Rosa: It sounds like they are providing a way to make those two traditionally separate problems work together in a unified, theoretically sound manner one <ref:2507.09134#pg1>.

Dev: And that unification is what makes it useful for real-time systems where we can't afford to solve massive OCPs constantly one <ref:2507.09134#pg1>.

Rosa: So, the main implication here is that we can get more reliable motion planning solutions in environments that are complicated and non-convex, without sacrificing the performance of a high-fidelity MPC controller one <ref:2507.09134#pg1>.

Dev: It moves the challenge from solving one massive optimization problem to solving a sequence of smaller, more manageable problems dictated by the path structure one <ref:2507.09134#pg1>.

Rosa: That’s a nice way to put it. We're trading one big hard problem for many smaller ones that are easier to solve sequentially one <ref:2507.09134#pg1>.

Title and authors: Dev: And that sequential nature is what makes it possible to maintain the necessary guarantees for stability and feasibility across the entire trajectory one <ref:2507.09134#pg1>.

Rosa: So, moving on, we discussed the paper's summary—what they actually propose in this Path Feasibility Governor framework.

Dev: They summarize it as a modular system that lets you integrate predictive controllers with a wide array of path planners two <ref:2507.09134#pg2,predictive controllers with a wide array of path planners>.

Rosa: And they show how given a target reference, the path planner generates a continuous path connecting your starting position to that target one <ref:2507.09134#pg1>.

Dev: Then the PathFG uses this generated path to create this sequence of intermediate references that are then passed directly into the MPC one <ref:2507.09134#pg1>.

Rosa: So, it’s a pipeline: planner makes a path, and the governor creates the steering signals along that path for the controller one <ref:2507.09134#pg1>.

Dev: This architecture is shown in their Figure one illustrating exactly how this integration works between the two main components two <ref:2507.09134#pg2>.

Rosa: And they emphasize that this approach is compatible with replanning, which means if something changes, you can just generate a new path and update those intermediate references one <ref:2507.09134#pg1>.

Dev: That modularity is important because it means you don't have to re-engineer the whole control architecture every time you need to adjust the trajectory in an environment with moving obstacles two <ref:2507.09134#pg2>.

Rosa: It sounds like they’ve made the process of replanning much more seamless and less computationally heavy than what was previously possible one <ref:2507.09134#pg1>.

Dev: And that brings us to how this translates into practical benefits, which is where we look at the improvements the authors suggest for this work.

Rosa: The improvements focus heavily on computational gains. They suggest that by using PathFG, you can achieve better performance by reducing the need for long prediction horizons five.

Dev: And they claim that this makes MPC controllers cheaper and more reliable because it decouples non-convex obstacle avoidance from simpler tracking problems five.

Rosa: That’s a big win for real-time applications. It means you can run a less demanding MPC solver with much faster iteration times five.

Title and authors: Dev: Then they discuss the region of attraction, which is another major point. They suggest that the system can achieve a significantly expanded region of attraction one <ref:2507.09134#pg1,a significantly expanded region of attraction>.

Rosa: They’re claiming this expands from just states reachable within the prediction horizon to any initial state that can actually connect to your target reference one <ref:2507.09134#pg1>.

Dev: And they back that up by saying it encompasses all initial states from which a feasible path can be generated, not just the immediately reachable ones one <ref:2507.09134#pg1>.

Rosa: That’s a huge safety margin. It means you have more room to maneuver before you hit an obstacle or run into a situation where no solution exists one <ref:2507.09134#pg1>.

Dev: So, they are essentially saying that the system's capability is much broader than what was previously proven for standard MPC setups one <ref:2507.09134#pg1>.

Rosa: They also point out the safety and stability guarantees. They proved recursive feasibility and safety for any target reference in a certain set, and asymptotic stability by showing an equilibrium point one <ref:2507.09134#pg1>.

Dev: That’s the theoretical underpinning. It shows that this isn't just a clever trick; it has mathematical proof behind its robustness one <ref:2507.09134#pg1>.

Rosa: And finally, they discuss the numerical validation. They showed real-time performance in quadrotor simulations in cluttered environments using a prediction horizon of five one <ref:2507.09134#pg1>.

Dev: That low prediction horizon, N equals five, is what really sells the real-time aspect because ungoverned MPC would need twenty just to stay feasible one <ref:2507.09134#pg1>.

Rosa: And they claim that the PathFG component alone maintains an average computation time under zero point one seconds, which is well below typical sampling periods one <ref:2507.09134#pg1>.

Dev: So, we’ve seen how they are pushing the envelope on both theoretical guarantees and practical computational speed in a real-world context one <ref:2507.09134#pg1>.

Rosa: This paper really shows how we can integrate planning and control more intelligently to get a system that is both safe and fast enough for deployment one <ref:2507.09134#pg1>.

Dev: It's a strong piece of work for anyone working on real-time trajectory generation in complex, dynamic environments one <ref:2507.09134#pg1>.

The paper's summary: Rosa: We’ve talked about the general improvements mentioned in this paper, and now we’re diving into the specific advantages they highlight. They suggest a few key areas where this PathFG framework really shines.

Dev: First up is computational efficiency, which is huge because it suggests reducing the need for long prediction horizons five.

Rosa: That's right. They argue that this makes MPC controllers cheaper and more reliable because it decouples the complex non-convex obstacle avoidance from the simpler tracking problems five.

Dev: So, instead of one giant OCP, you have a sequence of smaller ones that are easier to manage iteratively five.

Rosa: And then there’s the region of attraction expansion. They suggest that it moves from just the N-step backwards reachable set to any state connected to the target by a planner one <ref:2507.09134#pg1>.

Dev: That means if you're starting in a less ideal spot, you don't have to worry about being stuck because you can get there through a feasible path one <ref:2507.09134#pg1>.

Rosa: They claim this encompasses all initial states from which a feasible path P to the target reference can be generated, which is a big safety margin one <ref:2507.09134#pg1>.

Dev: That’s true. It expands the region of attraction beyond what was possible with standard MPC formulations one <ref:2507.09134#pg1>.

Rosa: And they mention how they handle non-convex motion planning. They suggest delegating those constraints to a path planner while letting the MPC deal only with convex constraints via affine over-approximations five.

Dev: That’s a smart way to approach it. It lets the MPC focus on what it's good at, which is handling those simpler, convex constraints reliably five.

Rosa: And they emphasize that this helps solve the overall optimal control problem more reliably and less expensively five.

Dev: So, for someone designing a system, this means you can rely on the MPC to handle its core tasks without getting bogged down in every single detail of the environment's geometry five.

Rosa: It’s about making the controller robust against environmental complexity by using that path structure as a guide one <ref:2507.09134#pg1>.

Dev: And finally, they stress that this leads to real-time performance. They show computation times well below the sampling period of zero point one seconds for both configurations one <ref:2507.09134#pg1>.

Rosa: So, in short, it’s a framework that delivers better safety and stability while keeping the computational load manageable for deployment one <ref:2507.09134#pg1>.

The paper's improvements: Dev: We’ve seen the general benefits now. Let's focus on the specific ways they claim this work improves upon existing methods, which is where we look at the detailed advantages.

Rosa: The first major improvement is definitely computational efficiency. They show that it reduces the need for long prediction horizons five.

Dev: And as we discussed, that directly translates to cheaper and more reliable MPC controllers by separating non-convex avoidance from tracking five.

Rosa: That’s because the PathFG handles the hard geometry problem externally, and the MPC only has to handle what it can solve well—the convex constraints five.

Dev: Then there's the region of attraction expansion. They claim this moves from just states reachable within a certain time window to any state that can be connected to your target by a planner one <ref:2507.09134#pg1>.

Rosa: That means you’re not restricted to being near the reference; you can start anywhere and still get there if a path exists one <ref:2507.09134#pg1>.

Dev: They emphasize that this encompasses all initial states from which a feasible path P to the target reference can be generated, which is a bigger safety margin than what was possible before one <ref:2507.09134#pg1>.

Rosa: And then they talk about how they handle non-convex planning. They use affine over-approximations of obstacles for the MPC and delegate the actual complex geometry handling to a path planner five.

Dev: That’s a pragmatic way to solve it. It lets you use simpler, more reliable convex constraints in your OCP formulation five.

Rosa: And they highlight that this makes solving the overall optimal control problem more reliable and less expensive five.

Dev: And finally, the real-time performance metric. They show computation times under zero point one seconds even with a short prediction horizon of five one <ref:2507.09134#pg1>.

Rosa: So, to put it all together for the listener, this paper delivers better safety guarantees and better speed by unifying planning and control in a way that's practical one <ref:2507.09134#pg1>.

Dev: It’s about making the controller more robust by using that path structure as a guide through complex spaces one <ref:2507.09134#pg1>.

Conclusion: Rosa: So, to wrap up our discussion on this paper, the Path Feasibility Governor. It really shows how we can integrate planning and control in a way that's practical for real-time applications.

Dev: It’s a strong contribution because it provides mathematical proof of safety and asymptotic stability for targets beyond the prediction horizon one <ref:2507.09134#pg1>.

Rosa: The main implication is that we can get better motion planning solutions in complicated, non-convex environments without sacrificing the performance of a high-fidelity MPC controller one <ref:2507.09134#pg1>.

Dev: It’s about making the controller more robust by using that path structure as a guide through complex spaces, and they provide concrete numbers on how much faster it is one <ref:2507.09134#pg1>.

Rosa: We should keep an eye on the limitations they mentioned, especially regarding real sensor noise and ensuring stability when the PathFG unit itself might have some issues one <ref:2507.09134#pg1>.

Dev: Right. And those are exactly the kinds of things we have to test when we move from simulation to real-world deployment, not just relying on perfect mathematical models one <ref:2507.09134#pg1>.

Rosa: This paper, "Integrating Planning and Predictive Control Using the Path Feasibility Governor," gives us a solid framework for building systems that are both safe and fast enough for deployment one <ref:2507.09134#pg1,Integrating Planning and Predictive Control Using the Path Feasibility Governor>.

More episodes

← Home