Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hybrid Imitation-Learning Motion Planner for Urban Driving

Cristian Gariboldi, Matteo Corno, Beng Jin

TL;DR

The paper tackles safe, human-like motion planning in urban driving by integrating imitation-learning with optimization. It introduces a hybrid architecture that uses a Planner to generate candidate paths, a Neural Network Trajectory predictor to imitate human driving, and a Model Predictive Trajectory module to enforce safety and feasibility, yielding a trajectory that balances fidelity and safety. The approach demonstrates improved closed-loop safety in both simulation and real-world deployment at low speeds, outperforming purely learning-based or optimization-based baselines in critical metrics and jerk smoothness. This work offers a practical path toward reliable urban autonomous driving by leveraging public datasets and real-vehicle validation.

Abstract

With the release of open source datasets such as nuPlan and Argoverse, the research around learning-based planners has spread a lot in the last years. Existing systems have shown excellent capabilities in imitating the human driver behaviour, but they struggle to guarantee safe closed-loop driving. Conversely, optimization-based planners offer greater security in short-term planning scenarios. To confront this challenge, in this paper we propose a novel hybrid motion planner that integrates both learning-based and optimization-based techniques. Initially, a multilayer perceptron (MLP) generates a human-like trajectory, which is then refined by an optimization-based component. This component not only minimizes tracking errors but also computes a trajectory that is both kinematically feasible and collision-free with obstacles and road boundaries. Our model effectively balances safety and human-likeness, mitigating the trade-off inherent in these objectives. We validate our approach through simulation experiments and further demonstrate its efficacy by deploying it in real-world self-driving vehicles.

Hybrid Imitation-Learning Motion Planner for Urban Driving

TL;DR

The paper tackles safe, human-like motion planning in urban driving by integrating imitation-learning with optimization. It introduces a hybrid architecture that uses a Planner to generate candidate paths, a Neural Network Trajectory predictor to imitate human driving, and a Model Predictive Trajectory module to enforce safety and feasibility, yielding a trajectory that balances fidelity and safety. The approach demonstrates improved closed-loop safety in both simulation and real-world deployment at low speeds, outperforming purely learning-based or optimization-based baselines in critical metrics and jerk smoothness. This work offers a practical path toward reliable urban autonomous driving by leveraging public datasets and real-vehicle validation.

Abstract

With the release of open source datasets such as nuPlan and Argoverse, the research around learning-based planners has spread a lot in the last years. Existing systems have shown excellent capabilities in imitating the human driver behaviour, but they struggle to guarantee safe closed-loop driving. Conversely, optimization-based planners offer greater security in short-term planning scenarios. To confront this challenge, in this paper we propose a novel hybrid motion planner that integrates both learning-based and optimization-based techniques. Initially, a multilayer perceptron (MLP) generates a human-like trajectory, which is then refined by an optimization-based component. This component not only minimizes tracking errors but also computes a trajectory that is both kinematically feasible and collision-free with obstacles and road boundaries. Our model effectively balances safety and human-likeness, mitigating the trade-off inherent in these objectives. We validate our approach through simulation experiments and further demonstrate its efficacy by deploying it in real-world self-driving vehicles.
Paper Structure (14 sections, 4 equations, 9 figures, 1 table)

This paper contains 14 sections, 4 equations, 9 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. Optimization-based planners
  4. Reinforcement Learning
  5. Imitation-Learning
  6. SYSTEM ARCHITECTURE
  7. Planner
  8. Multilayer Perceptron (MLP)
  9. Model Predictive Trajectory (MPT)
  10. EXPERIMENTS AND RESULTS
  11. Baselines
  12. ROS Simulator
  13. Real World Driving
  14. CONCLUSIONS

Figures (9)

  • Figure 1: Model's Structure | This is an overview of the hybrid model, showing the inputs and outputs of each block and the communication between interfaces. We also show the inputs provided by the Perception and Localization components such as observations c26, ego states and goal destination.
  • Figure 2: Four frames from self-driving simulations, the green line shows the neural network's output ("Neural Network Trajectory") and the pink line shows the optimization-based component's output ("MPT Trajectory"). The optimization process adjusts the neural network's output to avoid collisions with obstacles and road boundaries.
  • Figure 3: Left: default optimization-based planner's trajectory with velocity (top) and acceleration (bottom). Right: hybrid model trajectory with velocity (top) and acceleration (bottom). The vehicle's motion follows the arrow in the top left image. Blue lines indicate the lane boundaries.
  • Figure 4: Same as Fig. 3
  • Figure 5: Same as Fig. 3
  • ...and 4 more figures