Probability-Based Collision Risk Evaluation of Trajectories for Optimal Control Problems with Moving Obstacles

arXiv:2609.38534 · eess.SY, cs.SY · Submitted 2026-09-29 · Read on arXiv

Listen

Transcript

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

Rosa: Today's paper: "Probability-Based Collision Risk Evaluation of Trajectories for Optimal Control Problems with Moving Obstacles".

Dev: This paper presents a method for approximating occupancy distributions using smooth B-spline surfaces to enable time-dependent quantification of collision risk within optimal control problems.

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, we're diving into this paper now: "Probability-Based Collision Risk Evaluation of Trajectories for Optimal Control Problems with Moving Obstacles." The main idea seems to be using smooth B-spline surfaces to approximate occupancy distributions, which is crucial for making time-dependent collision risk quantification possible in optimal control problems.

Dev: Right, so it tackles the problem of planning collision-free paths for ground vehicles in dynamic settings by modeling the environment as a field with constant height that smoothly decays to zero at the edges. It claims this method allows for combining multiple occupancy distributions into one coherent risk representation that gives solvers gradient and higher-order derivative information they need.

Taro: That sounds important because standard potential field methods often use non-differentiable components, which limits what optimization algorithms can actually handle when you're trying to enforce smoothness or limit curvature. I wonder if this B-spline approach really solves that fundamental differentiability issue for trajectory planning?

Rosa: Exactly, Taro. The paper focuses on approximating the Poisson intensity function using these smooth surfaces to make sure the risk representation is differentiable, which is what allows the optimal control solvers to work properly with it. It also handles multiple obstacles by superimposing these fields instead of just dealing with them one by one.

Dev: From an engineering standpoint, that differentiability is key for the OCP formulation where you're minimizing things like control effort and keeping the trajectory smooth over time. I’m thinking about the loop rate here; if this approximation takes too long to compute, it won't help with real-time control at all.

Taro: The paper explores three different ways to interpret these combined occupancy distributions—the probability of collision-free traversal, Conditional Value-at-Risk CVaR, and an intensity-based interpretation using a spatio-temporal Poisson random field. Which one do you think is the most practical for handling unpredictable events when the world misbehaves?

Rosa: I'm leaning toward that intensity-based interpretation because it lets you aggregate multiple distributions without having to artificially constrain them into a simple zero one interval, which seems more flexible for complex scenarios. It feels like the most robust way to keep track of risk across various obstacle types.

Paper summary: Dev: But we have to think about the computational cost of that intensity function calculation; if it's too slow, we lose the benefit of having rich derivative information for trajectory selection. We need to ensure this fits within our required loop rate constraints.

Taro: And what about when things go seriously wrong? If the environment is highly uncertain and misbehaves, how does this framework adapt? Does it still give us a reliable plan even if the underlying occupancy distributions are highly inaccurate?

Rosa: The authors are trying to balance fidelity with smoothness through regularization techniques, specifically minimizing the Mean Square Error of the approximation while also penalizing high Total Variation and Tikhonov regularization terms. This regularization aims to keep the resulting surfaces smooth without completely losing the true underlying distribution information.

Dev: I've read about those metrics—the TV approximation penalizes large gradients, which helps suppress oscillations in the map representation, and Tikhonov regularization handles curvature by penalizing second derivatives. Those are essential for ensuring the resulting function is actually usable by derivative-based OCP solvers.

Taro: So it's a trade-off between how accurate the approximation is to the actual risk distribution and how smooth and differentiable we make that approximation, which speaks directly to those limitations mentioned in.

Rosa: Precisely, Taro. The paper acknowledges that methods relying on discretized state spaces need extra processing to generate these smooth trajectories, and this B-spline approach is one way to bridge that gap by providing continuous functions in a finite-dimensional subspace of the Sobolev space Wd,∞.

Dev: That means we're dealing with the complexity of solving a minimization problem for those B-spline coefficients Cˆ, which involves balancing MSE against TV and T terms. That minimization step is where latency could creep in if the system isn't designed carefully.

Taro: If we look at the application to OCPs, they use this intensity map λ(x, y;t) to formulate the objective function in their optimization problem. This suggests that not only does it inform where we might collide but also how much risk is associated with being in a certain spatial and temporal location.

Paper summary: Rosa: It really gives us more than just a binary collision check; it provides the gradient information needed for advanced trajectory planning. This level of detail lets the planner adjust its path proactively rather than just reacting to immediate threats from moving obstacles.

Dev: That proactive adjustment is what we want, but I'm still thinking about the practical deployment outside a perfect lab setting; how long can this system run reliably when the sensor data quality degrades or the environment changes rapidly?

Taro: That uncertainty in real-world conditions is precisely where these models are tested, and while they handle dynamic environments well, their performance depends heavily on how accurately those initial occupancy distributions are defined. If the input data is flawed, the output risk assessment will be too.

Rosa: That points to future work being important—we need methods that can handle noisy or incomplete sensor data better, ensuring the B-spline approximation doesn't just smooth over real errors. This paper lays a foundation for a more robust risk assessment tool.

Dev: So, to wrap up this overview of "Probability-Based Collision Risk Evaluation of Trajectories for Optimal Control Problems with Moving Obstacles," it seems the core value is providing a smooth, differentiable way to combine risk from multiple sources into an intensity function that directly feeds into optimal control solvers.

Taro: That’s right; the implications are in enabling trajectory planning that respects smoothness and incorporates complex risk metrics like CVaR or intensity fields in a way that is compatible with current optimization frameworks.

Rosa: And it gives us a clearer picture of the trade-offs involved when we choose between different risk representations, such as probability of safety versus CVaR.

Dev: The engineering challenge remains making that B-spline fitting process fast enough to support real-time operation while maintaining high fidelity for the control loop.

Taro: We need to keep pushing on how this framework performs when the world misbehaves and how it can be tuned for different risk tolerances in dynamic situations.

Rosa: It’s a solid piece of work that establishes a way to integrate occupancy distributions coherently, and we're excited to see where this methodology takes us in terms of deployment.

Conclusion: Rosa: So, we've seen how this paper uses smooth B-splines to make collision risk calculation time-dependent for optimal control problems with moving obstacles.

Dev: And that's because they are approximating occupancy distributions with these continuous surfaces, which gives us the necessary gradient information for the solvers.

Taro: I still wonder if we can trust this when things get really messy outside a controlled lab setting, Rosa.

Rosa: That’s a fair concern, Taro; we need to talk about what this approach actually achieves in practical scenarios.

Dev: From an engineering standpoint, my main worry is the computational load; how fast can the B-spline fitting process actually run without killing our loop rate?

Taro: And when the environment gets unpredictable—say, a sudden change in obstacle movement—how does this framework handle that uncertainty effectively?

Rosa: Well, it's about combining multiple risk views into one coherent representation, and that's what the authors are focusing on.

Dev: They investigate three ways to combine these distributions, like probability of safety or conditional value-at-risk, which tells us a lot about the interpretation.

Taro: The intensity-based interpretation seems interesting because it lets us work with occurrence rates rather than being stuck in strict probability bounds.

Rosa: Exactly; that flexibility is what makes it powerful for aggregating data from different sources in a dynamic scene.

Dev: But we have to be careful about that aggregation; if the input distributions are already noisy, the output intensity map could be misleading, which raises some issues for me regarding failure modes.

Taro: If we can use this intensity map to inform proactive planning, it means our vehicles won't just react to collisions but will actually adjust their paths based on predicted risk over time.

Rosa: That proactive adjustment is the whole point here; it moves us away from reactive navigation toward more intelligent, risk-aware pathfinding.

Dev: So, we're looking at a method that bridges the gap between complex probabilistic modeling and real-time control for autonomous systems.

Taro: And if they can manage those smoothness constraints effectively during the optimization process, it could significantly improve our autonomy in crowded environments.

Florian Steppicha, Matthias Gerdtsa

Institute of Applied Mathematics and Scientific Computing, Department of Aerospace Engineering, University of the Bundeswehr Munich

eess.SY, cs.SY

Submitted: 2026-09-29

Updated: 2026-09-29

Comments: 27 pages, 15 figures, presented 2024 on IFIP TC7, Hamburg, Germany

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 72/100

The gist: This paper presents a method for approximating occupancy distributions using smooth B-spline surfaces to enable time-dependent quantification of collision risk within optimal control problems.

Key concepts

Occupancy Distribution Approximation: The paper approximates the complex spatial occupancy distribution of obstacles using smooth B-spline surfaces. This mathematical technique ensures the resulting risk functions are differentiable, which is crucial for solving optimal control problems that rely on calculating gradients for trajectory optimization.
Risk Interpretation (CVaR): One key approach is using Conditional Value-at-Risk (CVaR). This measures the expected risk of outcomes in the worst scenarios, specifically focusing on the worst 100% minus a certain percentage ($\alpha$). It helps define a robust collision risk metric that accounts for high-impact, low-probability events.
B-Spline Surface Approximation: Occupancy distributions are modeled as smooth B-spline surfaces. These surfaces are defined by coefficients that are optimized to balance two competing goals: accurately matching the true occupancy data (Mean Square Error) and maintaining a smooth shape (Total Variation and Tikhonov Regularization).
Optimal Control Problem (OCP) Formulation: The final intensity map derived from the B-spline approximation is incorporated into an OCP. This allows the system to minimize a cost function that balances collision risk (weighted by the intensity map) against control effort, guiding the vehicle toward safe, collision-free paths.

Terminology

Regularization — 69× in this paper · explained in 3 episodes
Fidelity — 13× in this paper · explained in 2 episodes
Total Variation — 12× in this paper
Objective Function — 9× in this paper · explained in 2 episodes
Conditional Value-at-Risk — 5× in this paper
Mean Square Error — 4× in this paper
Steering — 4× in this paper · explained in 2 episodes
Accuracy — 3× in this paper · explained in 1 episode
Aggregation — 3× in this paper · explained in 1 episode
Adaptive Weighting — 2× in this paper · explained in 1 episode
Core — 2× in this paper · explained in 2 episodes
Higher-order derivatives — 2× in this paper · explained in 2 episodes
Motion Primitives — 2× in this paper · explained in 2 episodes
Multi-objective Optimization — 2× in this paper · explained in 4 episodes
Over-smoothing — 2× in this paper · explained in 4 episodes
Adaptation — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Consistency — 1× in this paper · explained in 3 episodes
Failure Modes — 1× in this paper · explained in 3 episodes
Model Predictive Control (MPC) — 1× in this paper · explained in 5 episodes
Temporal consistency — 1× in this paper · explained in 2 episodes
Trace — 1× in this paper · explained in 2 episodes

Summary

This paper presents a method for approximating occupancy distributions using smooth B-spline surfaces to enable time-dependent quantification of collision risk within optimal control problems. The proposed framework allows for the seamless integration of multiple occupancy distributions into a coherent risk representation, providing gradient and higher-order derivative information necessary for advanced trajectory planning in uncertain environments.

The gist: The method proposes approximating occupancy distributions using smooth B-spline surfaces to enable time-dependent quantification of collision risk within optimal control problems.

Modeling the Environment and Risk Representation

The method addresses the problem of planning collision-free trajectories for autonomous ground vehicles operating in dynamic environments populated by moving obstacles. The environment is modeled by constructing a time-varying map where each obstacle is approximated by a field with constant height that smoothly decays to zero at the edges, with heights scaled to reflect collision severity. Multiple obstacles are handled through superposition of these fields. The resulting risk function is intended to serve both as part of the objective function and constraints in optimal control problems (OCPs).

Interpretation of Collision Risk

The paper investigates three distinct interpretations for combining multiple occupancy distributions:

The probability of collision-free traversal, defined as the product of the complements of individual collision probabilities with each obstacle: Psafe(γ) = Y b∈B 1 − Z t∈T Z Z D(γ(t)) pb(x, yt) dx dy dt !.
Conditional Value-at-Risk (CVaR), which captures the expected risk of outcomes within the worst (1 − α) · 100% of collision scenarios: "CVaRα(γ) = E[CC > ν]."
The intensity-based interpretation using a spatio-temporal Poisson random field, where the occupancy distribution is reinterpreted as a spatial occurrence rate: λ(x, y;t) = X b∈B pb(x, yt). This formulation is favored because it allows for aggregation without the need to artificially constrain values to the interval [0, 1].

Approximation via B-Spline Surfaces

To ensure differentiability required by derivative-based OCP solvers, the Poisson intensity function is approximated using non-negative, continuous functions. The occupancy distributions are approximated by smooth B-spline surfaces, which span a finite-dimensional subspace of the Sobolev space Wd,∞. The approximation form is given as: pˆ(x, y;t) = nx+ X dx−1 i=1 ny+ X dy−1 j=1 cˆi,j (t)B dx i(x)B dy j(y), where the coefficients are collected in the spline coefficient tensor Cˆ.

Regularization and Optimization

The B-spline coefficients are computed by solving a minimization problem that balances fidelity and smoothness: min C ρMSE · MSE(˜p;t) + ρTV · TV(˜p;t) + ρT · T(˜p;t) s.t. ci,j (t) ≥ 0. The metrics used are:

Mean Square Error (MSE): Measures fidelity: MSE(˜p;t) = ∥p(x, yt) − p˜(x, yt)∥2 dx dy.
Total Variation (TV): Penalizes large gradients to suppress oscillations: TVapprox(˜p;t) = nx+ X dx−2 i=1 ny+ X dy−2 j=1 (ci+1,j (t) − ci,j (t))2 + 2 (ci+1,j+1 − ci+1,j−1 − ci−1,j+1 + ci−1,j−1)2 + (ci,j+1 − 2 ci,j + ci,j−1)2.
Tikhonov Regularization (T): Penalizes curvature to promote global smoothness: Tapprox(˜p;t) = nx+ X dx−2 i=2 ny+ X dy−2 j=2 (ci+1,j − 2 ci,j + ci−1,j) squared + 2 (ci+1,j+1 − ci+1,j−1 − ci−1,j+1 + ci−1,j−1)2 + (ci,j+1 − 2 ci,j + ci,j−1) squared.

Application to Optimal Control Problems

The resulting intensity map λ(x, y;t) is used in the OCP formulation: "min u Z t∈T Z Z D(γ(t)) λ(x, y;t) dx dy dt + ρu∥u∥2 2 s.t.

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed this paper, Probability-Based Collision Risk Evaluation of Trajectories for Optimal Control Problems with Moving Obstacles, which proposes a novel framework for integrating time-varying occupancy distributions into optimal control problems (OCPs) via B-spline surface approximation and intensity maps.

Here are the specific improvements to AI systems that can be made using this methodology, followed by what the improved system can achieve:

Specific Improvements to AI Systems:

Improve Trajectory Planning in Dynamic/Uncertain Environments via Risk-Aware Control:
Enhance Computational Efficiency of High-Dimensional OCP Solvers for Real-Time Decision Making:
Enable Smoother and More Robust Risk Assessments by Integrating Non-Linear Risk Measures (CVaR):
Create Flexible and Adaptive Risk Modeling by Utilizing Regularization Trade-offs (MSE vs. TV vs. Tikhonov):

What the Improved AI System Can Do:

Improve Trajectory Planning in Dynamic/Uncertain Environments via Risk-Aware Control:
The system can generate trajectories that are not just collision-free, but are optimized to minimize a quantifiable risk metric (like CVaR) rather than simply maximizing safety margins. This allows the vehicle to adopt more aggressive yet admissible maneuvers in dynamic scenarios, such as navigating narrow passages or anticipating high-risk obstacle clusters without halting unnecessarily.
Enhance Computational Efficiency of High-Dimensional OCP Solvers for Real-Time Decision Making:
By leveraging the B-spline intensity map formulation and regularization (especially the balanced approach), the system can significantly reduce solver computational load (up to 50% reduction reported) and improve convergence speed for high-dimensional, non-linear optimal control problems, enabling faster decision cycles crucial for real-time autonomous operation.
Enable Smoother and More Robust Risk Assessments by Integrating Non-Linear Risk Measures (CVaR):
The system can quantify the expected number of collisions in the worst (1 - α)·100% of scenarios, providing a more nuanced understanding of tail risks compared to simple probability calculations, allowing for risk acceptance strategies that are tailored to specific operational tolerances.
Create Flexible and Adaptive Risk Modeling by Utilizing Regularization Trade-offs (MSE vs. TV vs. Tikhonov):
The system can dynamically select the optimal approximation parameters (by tuning weights like ρMSE, ρTV, and ρT) based on the environmental conditions (e.g., using TC1 for smooth environments or TC3 for complex ones), allowing the AI to adapt its risk representation in real-time to balance fidelity against numerical stability and computational cost.

Abstract

This paper presents a method for approximating occupancy distributions, a common probabilistic representation used in trajectory planning for autonomous vehicles operating in uncertain environments. The proposed method employs B-spline surfaces in conjunction with regularization techniques to ensure smoothness and differentiability. Interpreting the resulting approximation as an intensity function of a Poisson random field enables seamless integration of multiple occupancy distributions into a coherent risk representation, while retaining compatibility with the semantics of probability distributions. The method is tailored to optimal control problems (OCPs), where solvers benefit from gradient and higher-order derivative information. It also preserves information about the spatial and temporal structure of occupancy distributions, which is critical for handling dynamic obstacles. We demonstrate the feasibility of the approach in a trajectory planning scenario involving autonomous vehicles and moving obstacles with time-varying uncertainty. By enabling controlled risk acceptance, our formulation extends beyond conservative "no-collision" strategies and allows for admissible trajectories that would otherwise be ruled out. In our application scenario, we report a reduction in solver computational load of up to 50% through proper tuning of the regularization.