MULTI-PROCESS 2D TRUCKING SIMULATOR USING MACHINE LEARNING ALGO-RITHMS AND INTELLIGENT AGENTS

Abstract

This paper describes the operation of compact, fast 2D simulator that provides Вird's-Eye View (BEV) and uses a baseline Proximal Policy Optimization (PPO) model to navigate an articulated tractor-trailer. The developed software system visualizes simple urban fragments and provides unified representation that supports collision checking, LiDAR, and rendering through single geometric pipeline.

The intelligent agent perceives a hybrid observation – structured numerical state plus a configurable LiDAR scan. The scan is computed by an efficient algorithm that scales nearly linearly with the number of LiDAR segments and rays. A minimal level of abstraction allows for easy switching between control types. Learning uses a small policy with two branches of the main simulator model (Multilayer Perceptron (MLP) for numerical characteristics and 1D Convolutional Neural Network (CNN) for LiDAR) in a vectorized, shared-memory configuration with program learning that incorporates accumulated experience and tightens the requirements for simulator performance as the corresponding software model becomes more sophisticated. Despite its deliberate simplicity, it does not impose restrictions on the main task considered in this paper (namely, the precise movement of articulated trucks at intersection level). The software demonstrates high learning success rates under tight tolerances on randomized scenes and supports interactive playback (using the keyboard or previously learned rules), making it a robust foundation and a convenient platform for rapid iteration (to improve simulator performance).

Further developments in this area, aimed at expanding the functionality of the proposed software package, will allow for scaling the variety of scenes, volume, and content of the corresponding learning program (due, for example, to an increase in the number of learning sequences and detailing of control without changing the architecture thanks to the unified architecture of the framework used).