This work proposes an a ordable setup for the application of machine learning to autonomous driving. We discuss the hardware details of the platform, and propose a scenario wherein learning from demonstration is leveraged to teach a vehicle how to drive. Preliminary results are discussed, together with potential targets enabled by the setup.
Autonomous vehicles have been attracting the attention of researchers in a
variety of engineering elds, and the rst successful attempts at navigating cars
autonomously can be traced back to more than thirty years ago | see, e.g., [
Given the current state of a airs, universities ought to introduce students to
Arti cial Intelligence (AI) themes that are useful in autonomous driving, as this
addition enhances their curricula and helps them to keep up with the state of the
art (see, e.g., [
More speci cally, our setup considers the problem of designing a controller for
the NXP-Cup challenge3, a competition which aims at introducing
undergraduate students to embedded programming through the solution of a (simpli ed)
autonomous driving problem for 1/18th scale battery-powered cars. Instead of
using traditional control approaches, i.e., manual synthesis and tuning of a
controller, we propose to synthesize a controller using machine learning techniques.
Our proposal is based on Learning from Demonstration (LfD) | see, e.g., [
Context. The setup we propose is meant to compete in the NXP-Cup, an annual context organized by NXP where students are required to build and program a model car, fully equipped with a range of sensors, so that it can run autonomously on a given track. The implementation of the car requires di erent skills, including embedded software programming, control theory, and some basic electronics.
Hardware. The basic hardware required for the competition is provided by NXP, which also dictates the rules teams must adhere to. The basic kit includes a battery-powered model car, a line scan camera used for sensing and a microcontroller to control the car. Additional sensors can be employed, such as wheel encoders or cameras, up to the limits de ned by the NXP rules. In our setup, the only additional sensor is a second line scan camera which enables a better reconstruction of the track unraveling in front of the vehicle. Given our proposal to teach the car how to drive using learning from demonstration, further components are required which will be then removed during the competition. In particular, we operate the car using a 2.4GHz hobby-grade remote controller coupled with a receiver. The signals from the controller are decoded to provide throttle and steering inputs, so that samples for our machine learning algorithm can be gathered under the form of reference trajectories. An Arduino board provides throttle and steering decoding, and stores demonstration data to be transferred 3 Previously known as Freescale cup https://community.nxp.com/docs/DOC-1284? tid=vanFREESCALECUP. coUmsmera’snd ROeNf./sOteFeFr Arduino steer
Line pos dataset
SD speed
DC motors KL25Z steer Steer pixels
Cams via Wi-Fi for further processing. In Figure 1 we show both a functional diagram of the components involved and a snapshot of the current prototype4. Control software. While most NXP-Cup competitors design and implement controllers manually, we propose to leverage state-of-the-art machine learning techniques. By doing so, we give students the chance to explore applications of AI solutions to a simple, yet challenging, autonomous driving problem. Students are required to formalize the learning problem as a LfD task and gather demonstration data by teleoperating the model car. After this initial phase, they have to choose the speci c learning technique | e.g., Neural Networks, SVRs, Gaussian Processes | and assemble the training dataset accordingly. This means they will have to reason and experiment about which sensory inputs are to be considered to obtain the best learning performances from a given learning algorithm. Ideally, students should be able to implement a simple state machine that includes the learned mapping from sensors to actuators in order to successfully drive the car along the track.
Simulator. A simulation of the model car running on a track has also been
implemented in Matlab. Using the simulator, students can focus on dataset and
algorithm design, avoiding hardware-related issues. A Simulink block containing
the kinematic model of the car has been implemented. This allows the user to
drive the car by specifying reference values for steering angle and speed. The
simulator shows the car displacing on a user-de ned track and also allows to
change the geometric parameters of the line-scan camera, obtaining di erent
eld of views, and therefore di erent performances. The code of the simulator
can be found in the companion site of this project. After data has been collected
from the simulator the learning process can take place. In our preliminary
implementation, demonstration data is fed to a Python script which implements
Gaussian Process Regression using the GPy library [
Preliminary experiments on the setup shown in Figure 1 have been carried out to investigate the feasibility of our proposal when it comes to teaching undergraduate students. In the following experiments, the car is not equipped with wheel encoders | which are a common addition in NXP competitions | and data about the steering angles is read directly from the motors. Other than this, the con guration of the car is NXP-Cup-legal, so it re ects intended usage scenarios. An instance of the LfD problem is considered where the objective is to learn a function mapping from camera readings at time t to steering angles to be applied at time t+1. Demonstration data was collected using the on-board architecture of Figure 1, learning was done o -board. Trials were run on two di erent tracks: (i.) a loop with two straights and two 180-degrees turns, and (ii.) a more complex track, including crossings of di erent shapes. The rst three authors, who are graduate students at the University of Genoa, tested the framework as if they were tackling the problem for the rst time. As a result, they managed to have the car learn how to drive on the rst track in a reliable way. However, the car could only move at a limited speed. The second track proved even more challenging: the car dealt badly with crossings, suggesting that improvements in sensory data and/or learning setup should be considered. While analyzing the dataset gathered during this experimental campaign we noticed that large parts of the input space where never explored during teaching. This was mostly due to sensors not being able to provide useful feedbacks for learning, e.g., line-scan cameras giving blank readings most of the time as the borders of the track were hardly seen during teaching. 4
We believe that, once some aspects related to reliable sensory acquisition are solved in the current prototype, the proposed setup will provide an excellent testbed for teaching undergraduate students basic principles in AI and Robotics. The hardware setup is a ordable and can be extended on a limited budget. Students can thus experiment easily how di erent sensors could a ect the outcome of the learning process. At the same time, the setup is standardized as the car must comply to the NXP-Cup rules. Furthermore, working on this problem allows to deal with a complete robotics project, from low-level control to high-level decision-making. On the AI side, the problem of teaching a car how to drive is not trivial, even in this simpli ed setup. Moreover, students have to deal with limited hardware resources, e.g., memory shortage, limited sensor feedbacks. On top of this, the controller obtained not only has to drive the car along the track, but it also has to be fast enough to compete with traditional controllers.