Ishaan Mahajan

Lower Fidelity Dynamic Models (LFDM) is a library of dynamics models for on-road wheeled vehicles that is written in C++ and CUDA and wrapped to Python using SWIG. Each model is described as a set of Ordinary Differential Equations (ODEs) that take a driver inputa normalized throttle between 0 and 1, a normalized steering between-1 and 1 (with-1 representing max steering toward a left turn), and a normalized braking input between 0 and 1, and subsequently advance the state of the vehicle (its position and velocity) forward in time.

This study explores the development of neural network-based control policies for autonomous robots, focusing on path following and obstacle avoidance. Utilizing the Autonomy Research Testbed (ART) and the Chrono simulation engine, we crafted two control strategies: an end-to-end imitation learning policy and a hybrid policy combining path following with a value function-based obstacle controller. Preliminary simulations validate both approaches, highlighting their respective efficiencies in managing complex navigation tasks. Future efforts will address transferring these policies to real vehicles, emphasizing the reduction of the sim-to-real performance gap.

We report on the transfer of an end-to-end control policy synthesized in simulation to a real-world set-ting. The policy guides a 1/6th scale vehicle, named ART-B, to a target location while navigating aroundobstacles with the aid of a 2D Lidar and GPS sensor. We utilize Gym-Chrono, a ReinforcementLearning (RL) environment based on the Project Chrono simulator, and the Open AI Gymnasium frame-work to synthesize this control policy trained using the Proximal Policy Optimization (PPO) algorithm.The approach involves training three versions of the policy: one for guiding ART-B across flat-rigid ter-rain, another for hilly-rigid terrain, and a third for hilly-deformable terrain. Subsequently, each policy will be tested in a real-world scenario with deformable terrain to answer the underlying research question– Does training an end-to-end control policy in a simulated setting with deformable terrain enhance itseffectiveness in real-world applications?

The SAE AutoDrive Challenge II is a four-year collegiate competition dedicated to developing a Level 4 autonomous vehicle by 2025. In January 2023, the participating teams each received a Chevy Bolt EUV. Within a span of five months, the second phase of the competition took place in Ann Arbor, MI. The authors of this contribution, who participated in this event as team Wisconsin Autonomous representing the University of Wisconsin–Madison, secured second place in static events and third place in dynamic events. This has been accomplished by reducing reliance on the actual vehicle platform and instead leveraging physical analogs and simulation. This paper outlines the software and hardware infrastructure of the competing vehicle, touching on issues pertaining sensors, hardware, and the software architecture employed on the autonomous vehicle. We discuss the LiDAR-camera fusion approach for object detection and the three-tier route planning and following systems. One of the defining aspects of our approach was leveraging early simulation and the use of physical analogs, which accelerated the development of the autonomy algorithms. In the process, we established a rapid autonomous vehicle development methodology that will anchor our technical effort in the third stage of the SAE AutoDrive Challenge II competition.

We describe a software framework and a hardware platform used in tandem for the design and analysis of robot autonomy algorithms in simulation and reality. The software, which is open source, containerized, and operating system (OS) independent, has three main components: a ROS 2 interface to a vehicle simulation framework (Chrono), which provides high-fidelity wheeled/tracked vehicle and sensor simulation; a basic ROS 2-based autonomy stack for algorithm design and testing; and, a development ecosystem which enables visualization, and hardware-in-the-loop experimentation in perception, state estimation, path planning, and controls. The accompanying hardware platform is a 1/6th scale vehicle augmented with reconfigurable mountings for computing, sensing, and tracking. Its purpose is to allow algorithms and sensor configurations to be physically tested and improved. Since this vehicle platform has a digital twin within the simulation environment, one can test and compare the same algorithms and autonomy stack in simulation and reality. This platform has been built with an eye towards characterizing and managing the simulation-toreality gap. Herein, we describe how this platform is set up, deployed, and used to improve autonomy for mobility applications.

This extended abstract reports on early results generated with a research testbed whose purpose is the characterization, measurement, and mitigation of the gap between simulation and reality in robots and off-road autonomous vehicles. Sometimes called the sim2real gap [1], it pertains to the difference in behavior of a robot whose autonomy stack is designed in simulation and then deployed in the real world. In most cases, the behavior in simulation is manifestly superior to the one in the real world. There are many causes for this, eg: the difference associated with the process of simulating sensors; the idealization of the dynamics model used to represent the motion of the real robot; shortcomings in capturing the world within which the robot operates; the non-deterministic temporal manner in which the underlying Robot Operating System (ROS) infrastructure operates, which leads to race conditions in producing and using information; and the difference in compute power available on an actual robot and when carrying out simulation. Ultimately, the sim2real gap is the observed inability of a control policy designed in simulation to work well with the real robot over a broad spectrum of operation regimes. Closing the sim2real gap is a case by case undertaking. In this abstract, we show a manifestation of this gap as it played out in conjunction with an in-house developed framework called ART/ATK–Autonomy Research Testbed/Autonomy Toolkit [2, 3], which is used to design and test autonomy stacks.