The Trajectory Planning Method for Intelligent Vehicles Using Reinforcement Learning and Risk-Reachable Sets

Safe RL and constrained optimization for structured-road trajectory planning.

Abstract

Trajectory planning is a key technology for ensuring the safety, comfort, and efficiency of autonomous vehicles. This study investigates the influence of spatially distributed driving risks on trajectory planning and proposes a framework combining risk-reachable sets, safe reinforcement learning, and constrained optimization.

First, a risk-reachable-set construction method is developed by integrating vehicle motion uncertainty with a driving-risk field. Gaussian-process regression is used to model surrounding-vehicle trajectories, while a probability-density-based risk assessment and Conditional Value at Risk are employed to identify high-risk boundaries over time and space.

Second, a safe reinforcement-learning method is proposed to generate initial trajectories. Polynomial trajectory parameters form the action space, and a constrained Markov decision process is solved using proximal policy optimization with a Lagrangian formulation. The generated trajectories are further projected onto the risk-reachable sets to improve safety and feasibility.

Third, a constrained iterative linear quadratic regulator is developed for trajectory optimization. Vehicle kinematics, comfort, efficiency, and safety are incorporated into the objective and constraints, while a barrier-function method ensures that trajectory points remain within safe regions.

Finally, a simulation platform is established, and high-speed merging experiments demonstrate that the proposed framework avoids high-risk areas while maintaining trajectory comfort, efficiency, and vehicle safety.

Keywords

  • Intelligent Vehicles
  • Trajectory Planning
  • Safe Reinforcement Learning
  • Risk-Reachable Sets
  • Constrained Iterative Linear Quadratic Regulator

Methods

Risk field setup, safe reinforcement-learning initialization, and constrained iterative LQR trajectory optimization workflow.
End-to-end planning framework: risk-field construction, safe reinforcement-learning initialization, and constrained iterative LQR refinement.

1. Risk Field and Risk-Reachable Set

A vehicle kinematic model and forward simulation produce nominal trajectories under Gaussian noise. Separate Gaussian-process regressors model lateral and longitudinal position uncertainty, which is converted into a time-aware driving-risk field. Forward reachability and CVaR then remove high-risk positions to form a safe, feasible risk-reachable set.

2. Safe RL Initial Trajectory

Trajectory planning is formulated as a constrained Markov decision process. PPO-Lagrangian learns a policy from reward and safety-cost signals, selects target positions, and produces longitudinal and lateral polynomial trajectories. A set-based projection corrects any trajectory point or speed that falls outside the risk-reachable set.

3. Constrained Trajectory Refinement

CILQR starts from the projected feasible trajectory and uses the risk-reachable set as a positional boundary. The objective balances jerk, steering-rate change, curvature, lateral offset, and safety distance; velocity, acceleration, steering, and obstacle constraints are handled with exponential barrier functions.

Experiments

Animated high-speed merging simulation with an ego vehicle and surrounding traffic.
High-speed merging simulation in the US-101 traffic scenario.

Compared with a Lattice planner in high-speed merging, the proposed planner generates safer and smoother trajectories.It also improves driving efficiency by reducing travel time and increasing average speed.

Results

Metric Lattice Baseline Proposed Planner Change
Average speed (m/s) 10.17 16.34 +60.7%
Maximum jerk (m/s3) 11.98 5.75 −52.0%
Maximum curvature (m−1) 0.02 0.01 −50.0%
Travel time (s) 7.40 4.80 −35.1%
Maximum risk 0.999 0.266 −73.4%
Average risk 0.417 0.074 −82.3%

The proposed planner increased average speed to 16.34 m/s while reducing travel time to 4.80 s. It also produced smoother trajectories, cutting maximum jerk by 52.0% and maximum curvature by 50.0%. Most importantly, the risk-aware reachable-set constraint reduced maximum risk by 73.4% and average risk by 82.3% relative to the Lattice baseline.