⚡ electrical engineering

Risk-aware and trackability-constrained trajectory planning for autonomous vehicles in unstructured road environments

This paper presents a comprehensive trajectory planning framework for autonomous vehicles in unstructured environments that integrates risk-aware path search, curvature-constrained optimization, and dynamic replanning to significantly reduce risk exposure and improve obstacle clearance while ensuring vehicle tracking feasibility.

Original authors: Xuan Yang, Enjun Bai, Jingsong Liu, Maochen Ge, Yongwang Chai, Shuang Chen, Wenyu Ma

Published 2026-10-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xuan Yang, Enjun Bai, Jingsong Liu, Maochen Ge, Yongwang Chai, Shuang Chen, Wenyu Ma

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Driving a car on a paved highway is a matter of following lines painted on the asphalt. The road tells the vehicle where to go, where to stay, and where the edges are. But drive off that pavement into a construction site, a forest trail, or a muddy field, and those lines vanish. The path is no longer defined; it is a chaotic landscape of uneven ground, hidden rocks, and unpredictable obstacles. For an autonomous vehicle, navigating this unstructured world is one of the hardest challenges in robotics. It must decide not just how to avoid a collision, but how to move safely through a space where the boundaries are fuzzy and the ground itself might be treacherous. The goal is to find a route that is not only free of crashes but also smooth enough for the car's wheels to follow without slipping or straining its steering mechanism.

Researchers at Shenyang Ligong University in China have developed a new way to help self-driving cars solve this problem. They created a planning system that acts like a cautious navigator, constantly weighing the dangers of the environment against the physical limits of the vehicle. Instead of simply looking for the shortest path between two points, their system calculates a "risk score" for every possible route. It understands that driving right next to a boulder is riskier than taking a slightly longer detour that keeps a safe distance, even if both routes technically avoid a crash. This approach, known as risk-aware planning, shifts the focus from merely surviving a trip to ensuring the journey is comfortable and secure.

The team tested their system on a digital map of a rough, unstructured environment filled with static obstacles like rocks and trees, as well as moving objects that could block the way. They compared their new method against a standard, traditional planning algorithm that simply looks for the shortest distance. The results showed that by prioritizing safety margins, their system successfully steered the virtual vehicle further away from obstacles. On average, the distance between the car and the center of nearby obstacles increased by 18.2 percent. At the same time, the total amount of risk the vehicle was exposed to during the trip dropped by nearly 20 percent. The trade-off was minimal: the path became only 1.5 percent longer, a small price to pay for a significantly safer route.

However, finding a safe path on a computer grid is only the first step. A route that looks safe on a map might still be impossible for a real car to drive. If the path requires the car to turn too sharply, the vehicle's steering system cannot physically follow it, or it might lose control. The researchers addressed this by adding a second layer of processing. After the initial route was found, they smoothed it out using a mathematical technique that bends the path into gentle curves, ensuring the turns were within the car's physical capabilities. This smoothing process reduced the sharpest turn the car would have to make by about 26 percent. Crucially, this smoothing did not push the car closer to obstacles; instead, it actually increased the minimum distance to the nearest obstacle by 12 percent, making the path both easier to drive and safer.

The system also had to handle the complication of moving obstacles. In the real world, a car might encounter a person walking across its path or a vehicle that stops unexpectedly. The researchers designed a hierarchical strategy to deal with these situations. If an obstacle is moving but will clear the path soon, the car simply slows down or speeds up to wait for the right moment to pass. This is a temporal adjustment, changing the timing of the drive without changing the route. But if an obstacle stops and permanently blocks the road, slowing down is not enough. In that case, the system triggers a local replanning routine, which calculates a completely new detour around the blockage. This dual approach allows the car to be patient when waiting is an option, but decisive when a new path is required.

To ensure the plan would work in reality, the researchers connected their software to a high-fidelity simulation of a real vehicle. They ran the planned trajectory through a closed-loop system where the car's computer tried to follow the path while the simulation tested the physics of the tires, suspension, and steering. The results were precise. The virtual car followed the planned route with a maximum side-to-side error of less than 7 centimeters and a heading error of less than 2.5 degrees. Most importantly, the steering wheel never reached its physical limit, meaning the car never had to turn harder than it was capable of. The system successfully balanced the need for safety, the constraints of the vehicle's mechanics, and the unpredictability of the environment.

While the study was conducted entirely in simulation, the findings offer a clear blueprint for how autonomous vehicles might navigate the messy, undefined spaces of the real world. The work demonstrates that safety is not just about avoiding collisions, but about managing risk and respecting the physical limits of the machine. By combining a risk-sensitive search with a smoothing process that respects vehicle dynamics, the researchers showed that a car can find a path that is both safe and drivable. The current prototype takes several minutes to compute a single route, which is too slow for real-time driving, but the framework proves that such a balance is achievable. Future work will focus on speeding up the calculations and testing the system on actual vehicles to see how it handles the noise and uncertainty of the physical world.

Technical Summary: Risk-aware and Trackability-constrained Trajectory Planning for Autonomous Vehicles in Unstructured Road Environments

Problem Statement
Autonomous driving in unstructured road environments presents significant challenges distinct from structured roads, primarily due to the absence of lane markings, irregular boundaries, and complex distributions of static and dynamic obstacles. Existing planning methods often face a trade-off: geometric path planners (like standard A*) may generate collision-free paths that hug obstacles or boundaries, offering insufficient safety margins for localization errors or model uncertainty. Furthermore, geometrically valid paths often violate vehicle kinematic constraints, such as maximum curvature, rendering them untrackable. Dynamic obstacles introduce additional complexity, where a spatially feasible path may become unsafe over time. Current approaches often treat risk assessment, path smoothing, and dynamic conflict resolution as separate problems, lacking a unified framework that balances cumulative environmental risk, vehicle trackability, and dynamic adaptability in unstructured spaces.

Methodology
The authors propose a unified, multi-stage trajectory planning framework designed to address these challenges through an integrated approach:

Risk-Aware Path Search (Enhanced A):*
The framework replaces the standard geometric cost function with a risk-sensitive cost. An enhanced A* algorithm incorporates a continuous risk field derived from obstacle proximity, road boundaries, and narrow passages using a quadratic distance-decay function. The cumulative path cost includes a motion cost (penalizing turns) and a risk term weighted by a coefficient ( $K_r$ ). An adaptive heuristic balances Euclidean distance, direction consistency, and Manhattan distance to guide the search. This approach prioritizes routes with lower cumulative risk exposure rather than merely avoiding binary collisions.
Trackability-Constrained Path Optimization (B-spline):
To ensure the generated path is physically executable, the discrete grid path undergoes B-spline optimization. The method represents the optimized path as a lateral displacement from the risk-aware reference path within a safety corridor defined by obstacle clearance and map boundaries. The optimization minimizes a cost function penalizing deviation from the reference, curvature energy, curvature variation, and path length, subject to hard constraints on maximum curvature (aligned with vehicle steering limits) and minimum obstacle clearance.
Hierarchical Dynamic Obstacle Handling:
The framework employs a hierarchical strategy for dynamic conflicts:
- Spatio-Temporal (S-T) Planning: For transient conflicts (e.g., moving obstacles), the system performs S-T speed planning. It uses Dynamic Programming (DP) followed by Quadratic Programming (QP) to generate a speed profile that adjusts acceleration, deceleration, or waiting to avoid collisions without altering the spatial path.
- Local Spatial Replanning (Hybrid A):* For persistent blockages (e.g., stopped obstacles) where temporal adjustment cannot restore feasibility, the system triggers local spatial replanning using Hybrid A*. This searches for a new geometric route around the blockage while respecting vehicle kinematics, followed by re-optimization of the speed profile.
Closed-Loop Validation:
The final reference trajectory is validated using a closed-loop simulation environment coupling a Simulink controller (lateral MPC and longitudinal cascade controller) with a CarSim vehicle model to verify tracking accuracy and steering saturation limits.

Key Results
Experiments were conducted on a single unstructured occupancy map with four different start-goal configurations, comparing the proposed method against standard A* and an enhanced A* without the risk term.

Risk Reduction: Compared to standard A*, the risk-aware search increased the mean obstacle-centre distance by 18.2% and reduced the risk integral and mean risk by 19.4% and 20.7%, respectively. This was achieved with only a 1.5% increase in average path length.
Trackability Improvement: The B-spline optimization stage successfully reduced the maximum path curvature from 0.2796 m⁻¹ to 0.2074 m⁻¹ (a 25.8% reduction), bringing it within the vehicle's tracking limits. Simultaneously, the minimum obstacle-centre distance increased by 12.1%.
Dynamic Conflict Resolution: Simulations demonstrated that transient conflicts were resolved via speed adjustments, while persistent blockages were successfully handled by local Hybrid A* replanning. The resulting trajectories maintained safe separation distances (e.g., >0.668 m for body-boundary distance) and satisfied all dynamic constraints (speed, acceleration, jerk).
Closed-Loop Tracking: In CarSim/Simulink simulations, the vehicle tracked the reference trajectory with a maximum lateral error of 0.0678 m and a maximum heading error of 2.470°. Crucially, no steering saturation occurred, confirming the compatibility of the planned curvature constraints with the vehicle's physical capabilities.

Significance and Claims
The paper claims that the proposed framework successfully links risk-aware route selection, curvature-constrained path optimization, and conflict-dependent replanning into a cohesive system for unstructured environments. The primary contribution is demonstrating that a unified approach can significantly reduce risk exposure and ensure vehicle trackability without incurring prohibitive path length penalties.

The authors modestly note that while the framework improves safety margins and trackability, it currently incurs higher computational costs (median search time increased from ~0.6s to ~1.5s, with total workflow time in the tens of minutes), indicating it is not yet a real-time solution. The study is limited to a single map topology and deterministic obstacle predictions. The authors emphasize that the results validate planning feasibility and closed-loop tracking in simulation but do not yet establish performance on physical vehicles or under perception uncertainty. Future work is identified as focusing on computational efficiency, probabilistic motion models, and hardware-in-the-loop testing.

Technical Summary: Risk-aware and Trackability-constrained Trajectory Planning for Autonomous Vehicles in Unstructured Road Environments

More like this