CARLA
Introduction to CARLA
CARLA: An open-source autonomous driving simulator.
Developed for the training, validation, and prototyping of urban driving systems.
Offers open digital assets including urban layouts, vehicles, buildings, and pedestrians.
Flexible specification of sensor suites and environmental conditions.
Overview of Research Approaches
CARLA is utilized to assess the performance of three approaches:
Classic modular pipeline: Vision-based perception module, rule-based planner, maneuver controller.
End-to-end model trained via imitation learning.
End-to-end model trained via reinforcement learning.
Evaluated across increasingly complex scenarios in an urban environment.
Performance metrics provided by CARLA demonstrate the utility of this platform.
Challenges in Autonomous Urban Driving
Complex environment with multi-agent dynamics (e.g., traffic intersections).
Necessity to track various actors (vehicles, pedestrians).
Recognition of road signs, lights, and other communication elements.
Need for quick decision-making in dynamic scenarios (e.g., pedestrians entering the road).
Risks associated with real-world testing, as many scenarios can be dangerous.
Benefits of Simulation
Simulation democratizes autonomous driving research,
Reduces high costs associated with physical world training.
Allows for extensive data collection in diverse scenarios without physical risks.
CARLA Features
Built on Unreal Engine 4; supports a dynamic world simulation.
Server-client architecture:
Server handles simulation rendering and environment dynamics.
Client interacts with the server to control the autonomous agent.
Environment Configuration
Contains 3D models of:
Static objects (e.g., buildings).
Dynamic entities like vehicles and pedestrians.
Two towns created for training (Town 1) and testing (Town 2).
Different atmospheric conditions available (lighting, weather).
Sensors and Readings
Configurable sensor suite in CARLA includes:
RGB cameras and pseudo-sensors for ground truth data.
Extensive customizability of camera parameters (type, location, orientation).
Approaches to Autonomous Driving in CARLA
1. Modular Pipeline
Divided into perception, planning, and control modules.
Perception Module: Uses semantic segmentation for environment mapping.
Local Planner: Generates waypoints based on environmental data.
Continuous Control: Uses a PID controller for vehicle dynamics.
2. Imitation Learning
Relies on recorded driving data from expert human drivers.
Dataset includes driving traces that consist of observations, commands, and actions.
High-level commands improve robustness of learned behavior.
3. Reinforcement Learning
Uses A3C (Asynchronous Advantage Actor-Critic) to optimize actions based on rewards.
Focuses on achieving navigation goals while avoiding obstacles.
Experimental Setup and Evaluation
Experiments conducted in controlled navigation tasks:
Tasks organized by difficulty (straight path, one turn, complex navigation).
Performance metrics: success rate in completing navigation tasks, inference analysis, collision data, etc.
Results indicate:
Modular pipeline and imitation learning generally perform well across tasks compared to reinforcement learning.
Generalization to unseen towns and weather conditions remains a challenge.
Conclusion
CARLA presents a powerful tool for autonomous vehicle research.
Promotes widespread experimentation and sharing of knowledge in the field.
Encourages further research on improving simulation-based training methods.