Towards the Design of Robotic Drivers for Full-Scale Self-Driving Racing Cars

Abstract

Autonomous vehicles are undergoing a rapid development thanks to advances in perception, planning and control methods and technologies achieved in the last two decades. Moreover, the lowering costs of sensors and computing platforms are attracting industrial entities, empowering the integration and development of innovative solutions for civilian use. Still, the development of autonomous racing cars has been confined mainly to laboratory studies and small to middle scale vehicles. This paper tackles the development of a planning and control framework for an electric full scale autonomous racing car, which is an absolute novelty in the literature, upon which we report our preliminary experiments and perspectives on future work. Our system leverages real time Nonlinear Model Predictive Control to track a pre-planned racing line. We describe the whole control system architecture including the mapping and localization methods employed.

LiDAR-Based GNSS Denied Localization for Autonomous Racing Cars

“…The frame of reference for q is taken within a pre-built occupancy grid map that is also used for LiDAR scan matching. The vehicle control system requires a 250 Hz pose estimate [ 36 ]: while the on board velocity and acceleration sensors are able to provide signals at this (or higher) frequency, the LiDAR scans are provided at 25 Hz. Hence, a smooth pose estimate at high frequency has to be carefully computed from sources with multiple rates.…”

“…Roborace’s DevBot 2.0 (London, UK), the car used during the race, has several sensors available, a subset of which is relevant to this work shown in Figure 2 . For an overview of the vehicle’s hardware architecture, see our previous work [ 36 ]. For the sake of this paper, the sensor data comes from the following three sources: OxTS Inertial Navigation System (INS): this commercial module consists of a dual-antenna GNSS and an IMU , which are pre-fused to obtain a high frequency (250 Hz) pose, velocity, and accelerations estimates; Ibeo LiDAR range finders : four LiDAR sensors are placed on the corners of the vehicle, each with 4 vertically stacked layers at 0.8 degrees spacing.…”

“…Some design choices were constrained by practical considerations related to the hardware available on the Roborace DevBot, specifically an NVIDIA DRIVE PX2 board with a non real-time Linux Kernel [ 36 ]. Thus, important constraints were in place in terms of: Computational burden: the NVIDIA Drive PX2 has a high number of CUDA cores, while CPU is rather limited; in this paper we propose a CPU implementation for simplicity and because the available computing power was enough for the particular experimental task; Flexibility: the particular race format affects not only strategy but also which sensors are available and what other modules must concurrently run on the board (e.g., interfaces with V2X race control infrastructure, planning software); Real-time requirements: the DevBot motion control module runs in real-time on a dedicated SpeedGoat board at 250 Hz.…”

“…With respect to our previous works, where an optimization based approach was used [ 36 , 37 ], we adopted a method based on particle filtering due to the lower computational burden required on the specific hardware and, in perspective, because it is more amenable to parallelization than optimization based approaches. The aforementioned drawback of this method (the kidnapping problem) was solved by injecting a priori knowledge of the particular scenario into the filter, as will be explained in detail in Section 5 , thus representing a valid alternative to optimization-based algorithms.…”

AMZ Driverless: The full autonomous racing system

“…In 2016, Roborace laid the groundwork for another type of full‐scale race competition with their DevBot platform. (Caporale et al, 2019) use this car to validate a nonlinear model predictive controller (MPC). Their work makes use of a preplanned racing line, whereas in this paper the racing line is generated online.…”