TL;DR
This paper introduces a novel autonomous racing strategy that dynamically switches between a learning-based controller for optimal lap times on an empty track and a parallelized optimization approach for safe and fast overtakes in the presence of other vehicles. The method achieves competitive lap times while ensuring collision-free maneuvers, even in complex multi-vehicle scenarios.
What Problem Does This Paper Address?
Autonomous racing presents a unique challenge: achieving the fastest possible lap times while simultaneously navigating a dynamic environment with other competing vehicles. Traditional autonomous driving approaches, focused on safety and adherence to traffic rules, are often too conservative for racing. Existing racing algorithms struggle to balance aggressive, time-optimal maneuvers with the safety requirements of avoiding other cars, especially when multiple vehicles are present. This often leads to suboptimal lap times or, worse, collisions and deadlocks. The paper aims to bridge this gap by providing a robust strategy that can handle both solo optimal performance and competitive multi-vehicle overtaking.
Methods
The core of the proposed strategy is a dual-mode planning and control system:
- Mode 1: No Surrounding Vehicles (Time Optimization): When the track is clear, a learning-based Model Predictive Control (MPC) trajectory planner is employed. This controller leverages historical lap data to iteratively refine its strategy, aiming for time-optimal performance by learning to push the vehicle to its dynamic limits. This is inspired by iterative learning control principles.
- Mode 2: Competing with Other Vehicles (Overtaking & Safety): When other vehicles are within the ego vehicle's overtaking range, the system switches to an optimization-based planner. This planner generates multiple "homotopic" trajectories in parallel. Each trajectory is designed to navigate around surrounding vehicles by following different lateral paths (e.g., between vehicles or between a vehicle and the track boundary). These reference paths are defined using Bezier curves.
- Parallelized Optimization: For each potential overtaking path, a separate optimization problem is solved to find a dynamically feasible and collision-free trajectory. This parallelization maintains low computational complexity even with varying numbers of surrounding vehicles.
- Safety Guarantee: A low-level safety-critical MPC controller tracks the selected optimal trajectory. This controller incorporates Control Barrier Functions (CBFs) to guarantee collision avoidance with surrounding vehicles, ensuring that the vehicle stays within a safe region.
The strategy uses a dynamic bicycle model for vehicle simulation and employs an affine time-invariant model for the trajectory planner and an affine time-varying model for the controller to manage computational load.
Datasets
Not discussed in the paper. The authors mention using a custom-designed simulator for their experiments, which includes state and input noises and uses information about surrounding vehicles' speeds and positions.
Key Insights and Ideas
- Dynamic Mode Switching is Crucial: The most significant insight is the necessity of a hybrid approach. A single strategy cannot optimally handle both solo performance driving and aggressive multi-vehicle racing. Switching between a time-optimal learning controller and a safety-aware parallel optimization planner is key.
- Parallel Homotopic Trajectory Generation: Generating multiple potential overtaking paths simultaneously (homotopic trajectories) is an effective way to explore different safe maneuvering options without significantly increasing computational burden. This allows the system to find the best compromise between speed and safety.
- Control Barrier Functions for Guaranteed Safety: The integration of CBFs in the low-level controller provides a strong guarantee of collision avoidance, which is paramount in competitive racing scenarios where aggressive maneuvers are required.
- Bezier Curves for Reference Paths: Using Bezier curves to define the homotopic reference paths simplifies the generation of smooth, dynamically feasible paths that can navigate around obstacles.
Key Results
- Competitive Lap Times: The proposed strategy allows the ego vehicle to achieve time-optimal performance when driving alone, comparable to learning-based MPC.
- Successful Overtakes: The system successfully demonstrated overtaking maneuvers in simulations with multiple randomly moving vehicles, both on straight and curvy track segments.
- High Success Rate: Overtaking success rates were high, especially for single and double leading vehicles, with rates of 100% and 98% respectively in simulations. Even with three leading vehicles, the success rate was 36%, with the strategy still guaranteeing safety.
- Steady Computational Complexity: The parallelized optimization approach maintained a low and steady computational complexity, with solver times around 39-40ms, enabling an update frequency of over 25 Hz.
- Deadlock Avoidance: The parallel generation of multiple trajectory options helps the ego vehicle avoid deadlocks, a common issue in complex overtaking scenarios.
Limitations and Future Work
- No Interaction Modeling: The current simulation does not model the interaction between the ego vehicle and other surrounding vehicles beyond their predicted trajectories. This means the simulator doesn't account for how the ego vehicle's actions might influence the other cars' behavior.
- Future Work: The authors explicitly mention considering "interaction between the ego vehicle and other surrounding vehicles" in future work. They also suggest exploring "blocking cars from overtaking" as a potential future strategy.
Summary updated