Computer Vision & Graphics Machine Learning & Data Science Research

PGDrive Simulator Generates Unlimited Diverse Driving Environments

Researchers proposed PGDrive, a driving simulator designed to evaluate and improve end-to-end driving agents' generalization abilities.

Researchers from The Chinese University of Hong Kong, SenseTime Research and Zhejiang University have proposed PGDrive, a driving simulator designed to evaluate and improve end-to-end driving agents’ generalization abilities.

Driving simulators such as CARLA and SUMMIT are currently widely used for end-to-end training of autonomous driving systems. Although such simulators feature visually realistic driving depictions, they typically include only a fixed set of maps and a limited number of configurations. Models trained on a small number of examples often struggle with new scenarios.

“It is known that deep neural networks can overfit training data easily,” explain the researchers. PGDrive, meanwhile, is able to generate an unlimited number of diverse driving maps. The simulator is introduced in the paper Improving the Generalization of End-to-End Driving through Procedural Generation.

image.png

The researchers identify a key PGDrive mechanism as procedural generation (PG), which enables a diverse range of executable maps to be procedurally generated from predefined, elementary “road blocks.” The video game industry first adopted PG algorithms to automatically generate diverse game content such as different levels, maps, racing courses, etc. Machine learning researchers have recently picked up on the benefits of using PG algorithms to generate training samples with diverse training settings to tackle the challenge of data overfitting.

The new paper introduces the concept of road blocks — elementary roadway components that can be assembled into complete maps by the procedural generation algorithm.

The PGDrive model builds its maps using seven road block types, each with configurable settings:

  • Straight: A straight road can be configured by the number of lanes, lane length and lane width, as well as by lane line type (broken or solid white line).
  • Ramp: A road with entry or exit existing in the rightest lane. Acceleration lane and deceleration lane are attached to the main road to guide vehicles to their destination.
  • Fork: A structure used to merge or split additional lanes from the main road and change the number of lanes.
  • Roundabout: A circular junction with four exits (sockets), whose radius is configurable. Roundabout, ramp and fork can be configured to provide diverse merge scenarios.
  • Curve: A curve block consists of a curved road and a straight road. The curve can be configured as a simple circular shape or a clothoid (spiral) shape to smooth the curvature variation.
  • T-Intersection: An intersection that can be entered and exited in three ways (three sockets). The turning radius is configurable.
  • Intersection: A four-way intersection that allows bidirectional traffic.
image.png
image.png
image.png

Under predefined rules, the building blocks are randomly selected by the PG algorithm and assembled into environments where the driving agents can interact. The team says that while driving agents such as Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) trained on a small fixed set of maps generalize poorly to unseen maps, their generalization ability across scenarios with different traffic densities and map structures improves as their training environment grows with more procedurally generated maps.

image.png

The researchers have made the PGDrive code available on their project GitHub, and the paper Improving the Generalization of End-to-End Driving through Procedural Generation is on arXiv.


Reporter: Fangyu Cai | Editor: Michael Sarazen


B4.png

Synced Report | A Survey of China’s Artificial Intelligence Solutions in Response to the COVID-19 Pandemic — 87 Case Studies from 700+ AI Vendors

This report offers a look at how China has leveraged artificial intelligence technologies in the battle against COVID-19. It is also available on Amazon KindleAlong with this report, we also introduced a database covering additional 1428 artificial intelligence solutions from 12 pandemic scenarios.

Click here to find more reports from us.


AI Weekly.png

We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.

4 comments on “PGDrive Simulator Generates Unlimited Diverse Driving Environments

  1. Pingback: [R] PGDrive Simulator Generates Unlimited Diverse Driving Environments – tensor.io

  2. Pingback: PGDrive Simulator Generates Unlimited Diverse Driving Environments – NikolaNews

  3. Pingback: [R] PGDrive Simulator Generates Unlimited Diverse Driving Environments – ONEO AI

  4. very good thanks

Leave a Reply

Your email address will not be published.

%d bloggers like this: