Brazilian university tests self-driving taxi service

Driverless vehicle can be called using smartphone app, which allows user to use voice commands to say where to go.

By Elton Alisson

Agência FAPESP – Anyone who happens to be on the São Carlos campus of the University of São Paulo (USP) in Brazil who sees a taxi moving along the road with no driver, but with passengers, can rest assured that it is not a runaway vehicle.

A taxi service using an autonomous vehicle is being tested by researchers at USP’s Mathematics & Computer Science Institute (ICMC) and the same university’s São Carlos School of Engineering.

The researchers are completing the final tests and adjusting a few details of the vehicle in preparation for a public demonstration of the autonomous taxi service, scheduled for mid-October.

“The idea is for a driverless taxi to be called using a smartphone app we’re also developing,” said Denis Wolf, a professor at ICMC-USP and coordinator of the project. “The taxi will go anywhere on campus. Customers will input the destination by voice command or using a touch screen inside the car. The taxi will then return to its parking spot to await the next call,” Wolf told Agência FAPESP.

The autonomous taxi service is one of the possible applications imagined by the researchers for CARINA, the Intelligent Robotic Car for Autonomous Navigation (Carro Robótico Inteligente para Navegação Autônoma, in Portuguese) that they have developed over the past few years with support from FAPESP and the National Scientific & Technological Development Council (CNPq) under the aegis of the National Science & Technology Institute for Critical Embedded Systems (INCT-SEC).

CARINA is only one of the autonomous cars under development in Brazil. Others are in the pipeline at the Federal Universities of Minas Gerais (UMFG) and Espírito Santo (UFES). However, it is the first in Latin America to have been tested on real city streets. CARINA’s test drive took place in early October 2013, when it successfully self-navigated for 5.5 km in São Carlos (read more at http://revistapesquisa.fapesp.br/en/2014/01/31/driverless-car).

Since then, the car, a Fiat Palio Weekend Adventure purchased from a dealership and adapted by the researchers with special equipment, has undergone several enhancements.

One of these was the installation of a continuous mapping system that improves vehicle control and localization as well as route planning, Wolf explained. “Route planning for CARINA’s first test run in October 2013 was very simple, based on a GPS system equivalent to sat-nav. The continuous mapping to be used in the next test will enable real-time planning of the car’s route to the destination input by the passenger,” he said.

Localization system

According to Wolf, self-driving vehicles localize themselves using a combination of information from GPS sensors and metric maps of the areas in which they are designed to circulate. These maps are specially constructed to pinpoint the vehicles’ street location.

However, even the most sophisticated GPS sensors are subject to failure and are relatively imprecise, especially on urban streets, owing to the presence of tall buildings and trees. Localization errors are often significant as a result, and loss of the GPS signal for even a few seconds prevents route correction.

The metric maps used today are grid based and vulnerable to service interruptions, consuming a large amount of memory.

As part of an ongoing project funded by FAPESP, the researchers plan to replace metric maps with continuous maps developed at the University of Sydney in Australia, duly adapted to enable CARINA to self-localize in São Carlos.

Continuous maps use different kinds of additional information on the environment, such as lane markers, traffic signs, curb height and position, and other features, to aid vehicle localization in vertical and horizontal space. They plot continuous functions between topological spaces, without the grids used in metric maps, and are far less sensitive to signal interruptions, Wolf explained.

“Continuous maps permit autonomous vehicle localization based on information like curb height and position, so you don’t depend on GPS sensors,” he said.

CARINA maps the area in which it travels by means of two laser sensors, placed at the front and on the roof of the vehicle, as well as 360-degree cameras similar to those used by Google Street View. The vehicle self-localizes based on previously filmed locations featuring details of the surrounding terrain.

Two laser sensors also scan 360 degrees and collect 700,000 data points per second, mapping everything around the car in a radius of 50 m and measuring the distance to other cars, poles, people, dogs and curbs, among other objects or obstacles, as well as their angle and height relative to the vehicle.

A twin-lens stereo vision system operates in tandem with the laser sensor installed on the front of the car, gauging the depth of objects around the vehicle and interpreting the position of traffic lanes.

“This set of sensors enables the autonomous vehicle to ‘see’ nearby vehicles, measuring their speed and direction so as to avoid the risk of collision,” Wolf said. The researchers have also developed a speed and steering control system based on command software. The car’s maximum speed is currently 60 kph.

“The maximum permissible steering control error for an autonomous vehicle is 40 cm. More than that risks putting it on the wrong side of the road and causing a collision with oncoming traffic,” Wolf said.

Autonomous truck

After they developed CARINA, the São Carlos group of researchers was asked by Scania AB to develop an autonomous truck.

The Swedish truck and bus maker provided two trucks for the project, in which it invested R$1.2 million (now about US$344,000).

Various devices were fitted to the truck so that the autonomous system can control every movement, including small motors to control the steering wheel and brakes. Speed is controlled by an electronic circuit attached to the accelerator.

A computer connected to all systems in the truck is responsible for capturing GPS and other sensor data, interpreting it, and selecting the appropriate command for actions such as accelerating, steering or braking.

“The solutions we used in the autonomous truck differed from those used in CARINA owing to differences between the platforms and budgetary constraints,” Wolf said. “We decided to do without the laser sensors used in CARINA, for example: they cost up to twice the price of the car. Instead, we used radar to detect obstacles and a pair of stereo cameras at the front of the truck.”

The cameras used by the researchers mimic human vision by capturing two images at a time so that the depth and shape of objects such as traffic lights can be calculated.

GPS antennas were installed on top of the cab, and a sensor detects movements of the steering column.

According to Wolf, one of the features that distinguishes the autonomous truck developed by USP’s São Carlos group from others produced worldwide, such as the Mercedes-Benz Future Truck introduced by Daimler, is the use of continuous maps like those used in CARINA.

“Continuous mapping enables the truck we developed to operate on dirt roads or in other situations where there are no lane markings and even in the absence of any traffic signs,” he said. “In this sense, our project is better suited to local conditions.”

The autonomous truck is a prototype and circulates only in a limited area of the São Carlos campus, but the results obtained so far are promising, according to Wolf.