The development of an autonomous driving stack (ADS) is challenging since it requires rigorous testing at each step. Whether the stack is modular, semi-modular, or end-to-end, the testing pipeline always follows a simulation to real-world testing hierarchy before any new feature is deployed on the autonomous vehicle (AV). On an abstract scale, it is difficult to keep track of what improvements have been made to the stack and what changes degraded the overall performance of the autonomy. This problem worsens when multiple research teams are actively contributing to the development. In this paper, we discuss a scenario-driven development approach that guides the development of the autonomy stack. We use Autoware Mini, an internally developed modular autonomy stack (currently in its early stages of development), as a case study and evaluate its performance as development progresses with Software-In-loop (SIL) and holistic testing. We use CARLA Leaderboard and Scenario Runner to evaluate the stack with automated benchmarking and discuss the evaluation feedback, specifically w.r.t. perception and prediction module of the autonomy stack.

Overtaking vehicles is a risky manoeuvre for human drivers and an even more difficult challenge for autonomous cars. The algorithms for overtaking require extensive information about the surrounding environment including knowing the length of a leading vehicle. The usual sensing modalities used in autonomous vehicles (vision, radar, LiDAR) are not suitable for estimating that length. In literature, flow sensing has been shown to aid underwater robots in navigation and localization. This suggests that flow sensing could also provide useful information for autonomous vehicles. This study investigates air flow data behind truck-sized bluff bodies using data acquired from Computational Fluid Dynamics (CFD) simulations. The proposed features for classification are based on Fast-Fourier transforms. The results show that pressure data can be used to differentiate between various truck lengths, indicating that flow sensors could aid autonomous vehicles in overtaking.

Autonomous driving is of great interest to industry and academia alike. The use of machine learning approaches for autonomous driving has long been studied, but mostly in the context of perception. In this article, we take a deeper look on the so-called end-to-end approaches for autonomous driving, where the entire driving pipeline is replaced with a single neural network. We review the learning methods, input and output modalities, network architectures, and evaluation schemes in end-to-end driving literature. Interpretability and safety are discussed separately, as they remain challenging for this approach. Beyond providing a comprehensive overview of existing methods, we conclude the review with an architecture that combines the most promising elements of the end-to-end autonomous driving systems.

Flow sensing has been investigated in the context of underwater and aerial robotics in the past decade. It has not been explored for applications in autonomous ground robotics. In this work-in-progress paper, we investigate the use of air-flow sensing for the applications in autonomous driving. More precisely, we investigate the use of air-flow sensing for vehicle length estimation by conducting computational-fluid-dynamics (CFD) simulations.

AGV (automated guided vehicle) systems are extensively used in factory and warehouse environments. As most of these environments employ a mix of AGVs, manually driven vehicles, and human workers, safety is an important subject. Current AGV systems employ safety fields and laser scanners to ensure safety in their environments. These fields however are often primitive and do not take into account future AGV trajectory or intentions of agents in their vicinity. This results in inefficient operation of such AGVs. We propose a three-layered architecture that consists of safety fields that are formed around immediate future trajectory of AGV as well as on the predicted intention of an agent in AGV vicinity, resulting in more efficient AGV behaviour. Results are presented using real laser data from a small-sized lab AGV as well as an industrial forklift truck.

Underwater robots conventionally use vision and sonar sensors for perception purposes, but recently bio-inspired sensors that can sense flow have been developed. In literature, flow sensing has been shown to provide useful information about an underwater object and its surroundings. In the light of this, we develop an underwater landmark recognition technique which is based on the extraction and comparison of compact flow features. The proposed features are based on frequency spectrum of a pressure signal acquired by a piezo-resistive sensor. We report experiments in semi-natural (human-made flume with obstacles) and natural (river) underwater conditions where the proposed technique successfully recognizes previously visited locations.

This paper presents a technique for the calibration of multi-beam laser scanners. The technique is based on an optimization process, which gives precise estimation of calibration parameters starting from an initial estimate. The optimization process is based on the comparison of scan data with the ground truth environment. Detailed account of the optimization process and suitability analysis of optimization objective function is described, and results are provided to show the efficacy of calibration technique.