Learning Homotopy Prediction for Optimization-Based Trajectory Planners for Autonomous Driving

Abi Rahman Syamil, Dongsuk KumInternational Journal of Automotive Technology 2025 · DOI

Abstract

Towards the real-world deployment of autonomous vehicles, it is crucial that autonomous driving systems plan safe, collision-free trajectories. However, generating a collision-free trajectory around potential obstacles requires solving a non-convex problem for conventional optimization-based planners. This is characterized by multiple local minima reflecting possible maneuvers in different homotopies that the ego vehicle can execute to avoid collision. Finding a solution among these maneuvers involves combinatorial decision-making, which incurs increasing computational costs as more obstacles are considered. In this paper, we propose a hybrid approach using learning-based decision-making for optimization planning. Our learning-based decision-maker predicts homotopic boundary constraints for the optimization planner, effectively determining a maneuver for the ego vehicle without any combinatorial decision process. In addition, the homotopic bounds enable us to reformulate the non-convex optimization problem into a more tractable quadratic programming (QP) problem. We evaluate our approach in unsignalized intersection scenarios using a simulator, demonstrating that it achieves better driving performance than existing decision-making and planning methods in non-convex driving situations.

Introduction

Planning safe paths for autonomous vehicles (AVs) in complex environments like unsignalized intersections is a significant challenge. The primary difficulty lies in the non-convex nature of the solution space. For instance, when approaching an intersection with an oncoming vehicle, the AV must decide whether to accelerate and pass before it or decelerate and pass after it. These distinct topological choices are known as homotopies.

Illustration of different homotopies and their corresponding path-time boundary constraints for an ego vehicle.

Traditional optimization-based planners, such as Model Predictive Control (MPC), often struggle with these scenarios. They may get trapped in local minima or require exhaustive combinatorial searches to evaluate every possible passing sequence. As traffic density increases, this "combinatorial explosion" leads to high computational costs, often resulting in "freezing" behavior where the vehicle fails to find a timely solution.

Methodology: Hybrid Trajectory Planning

To address these issues, this research proposes a hybrid framework that combines the strengths of deep learning for high-level decision-making with the precision of classical optimization for trajectory generation.

The proposed hybrid framework, integrating a Homotopy Prediction network with an Optimization Planner to generate safe trajectories.

Homotopy Prediction

The "brain" of the system is a transformer-based neural network designed to predict homotopic constraints. Instead of predicting a full trajectory, the network classifies each surrounding obstacle into one of three categories relative to the ego vehicle:

  • Upper Bound: The ego vehicle should pass after the obstacle.
  • Lower Bound: The ego vehicle should pass before the obstacle.
  • Ignore: The obstacle does not significantly constrain the ego vehicle's path.

The architecture utilizes an Ego Encoder and an Agent Encoder to process state data, followed by a Transformer Decoder. The attention mechanism within the transformer allows the model to effectively capture the complex interactions between multiple vehicles simultaneously.

Optimization Planner

Once the homotopy is predicted, the high-level decision is converted into boundary constraints. This transforms the original non-convex optimization problem into a Quadratic Programming (QP) problem. Because the combinatorial decision has already been made by the neural network, the QP solver can efficiently find the smoothest and most efficient path within the defined boundaries. This approach is mathematically tractable and ensures the final trajectory adheres to the physical limits of the vehicle.

Experimental Evaluation

The framework was evaluated using the SUMO (Simulation of Urban MObility) simulator in a four-way unsignalized intersection. The dataset comprised approximately 30,000 frames of expert driving data generated by an A* planner.

Key Results

The proposed method was compared against three baselines: an A Planner, a Time-To-Collision (TTC) rule-based planner, and a Behavior Cloning (BC) model. The proposed hybrid approach achieved a 95.5% success rate, outperforming A (92.5%), BC (79.0%), and TTC (74.5%).

In high-traffic scenarios, such as complex left turns, the proposed method maintained an 88% success rate, significantly outperforming the TTC planner (38%). Notably, the hybrid approach achieved a higher success rate than the A* "expert" it was trained on, as it generalized better and avoided the timeouts common in conservative search-based methods. Safety was also significantly improved; while the BC model had a 21% collision rate, the proposed method maintained a much lower 4% collision rate.

Conclusion and Future Work

By predicting constraints rather than direct paths, this hybrid approach maintains the safety and smoothness of optimization-based planning while gaining the speed and decision-making capabilities of deep learning. The model's outputs are also interpretable, allowing developers to understand the reasoning behind specific maneuvers.

Current limitations include a focus on longitudinal planning and the assumption of constant acceleration for other agents. Future work will explore lateral homotopy (e.g., overtaking decisions) and validation on real-world datasets to account for reactive agent behavior.

Summary updated

Abstract

Abstract Towards the real-world deployment of autonomous vehicles, it is crucial that autonomous driving systems plan safe, collision-free trajectories. However, generating a collision-free trajectory around potential obstacles requires solving a non-convex problem for conventional optimization-based planners. This is characterized by multiple local minima reflecting possible maneuvers in different homotopies that the ego vehicle can execute to avoid collision. Finding a solution among these maneuvers involves combinatorial decision-making, which incurs increasing computational costs as more obstacles are considered. In this paper, we propose a hybrid approach using learning-based decision-making for optimization planning. Our learning-based decision-maker predicts homotopic boundary constraints for the optimization planner, effectively determining a maneuver for the ego vehicle without any combinatorial decision process. In addition, the homotopic bounds enable us to reformulate the non-convex optimization problem into a more tractable quadratic programming (QP) problem. We evaluate our approach in unsignalized intersection scenarios using a simulator, demonstrating that it achieves better driving performance than existing decision-making and planning methods in non-convex driving situations.
BibTeX
@article{syamil2025learning,
  author = {Abi Rahman Syamil and Dongsuk Kum},
  title = {{Learning Homotopy Prediction for Optimization-Based Trajectory Planners for Autonomous Driving}},
  journal = {International Journal of Automotive Technology},
  year = {2025},
  doi = {10.1007/s12239-025-00257-1},
}
Figure