czech english

Berta

vacuum cleaner controlled by microcontroller

Berta is the winner of the Cleaning contest 2002! It is the prove that the simplest solutions are often the best. It has beaten its competitors with 2D range finders, cameras and a lot of processing power simply because it worked the way it was supposed to using only a front bumper and a microcontroller. Find out how eXtreme programming concepts work while building a robot.

Quick Info

The boggie is made of aluminum U-shaped profiles, powered by two DC motors (1.5Amp/12VDC, 5500rpm with 20:1 gearbox). The brain of the robot is 16bit Hitachi microcotroller (14.745 MHz clock, 4kB RAM, 128kB ROM (flash)). The only sensor that this robot utilizes is its front bumper. The cleaning method is based on a random walk algorithm.

Mechanical Design

The boggie is made of U-shaped aluminum profiles. Aluminum has proven itself to be a relatively light and quite durable material for such a construction. The bogie is a triangle with two geared motors mounted in the base, which is also front of the robot. The opposite vertex provides support for a caster wheel. The motors are 1.5Amp/12VDC, 5500rpm with 20:1 gearbox (dimensions: 60 x 33 x 33mm).
The cleaning unit is a plastic box made with a hot-glue. The box is removable and it is attached to the robot by aluminum holder. Dry vacuuming is used for collection of dirt. The cleaning mechanism consists of 2 salvaged 12V car vacuum cleaners and a box for collecting the dirt, equipped with 2 nozzles. An additional brush helps to collect dirt before it is sucked into the box.
The power source for all motors and electronics is 12V 7AH/20AH sealed rechargeable lead battery.
Both motors are powered by dual H-bridge and controlled via PWM (Pulse Width Modulation) directly from the main microcontroller. A feedback is provided by two encoders taken from a regular computer mouse. A non-linear PID controller is used for maintaining speed and direction.

Onboard Computer

Single microcontroller Hitachi H8/3048 - 14.745 MHz clock, 4kB RAM, 128kB ROM (flash). During software development we have found, that it is more convenient to drive robot as a slave from a PC through a serial interface. The H8 processor was programmed to fulfill the simplest tasks only and the program itself run in the PC. The PC allowed precise logging and easier changes in program. The same program was programmed to the H8 chip only after the software development was finished.

Sensor System

There are only two means of sensing the environment for the robot:
  • bumper micro-switches
  • virtual bumper based on correlation between applied power to the motors and resulting speed measured by encoders
The switches are very reliable and do not need any filtering. There are scanned every millisecond in the main processing loop where the change of state can be directly triggered.
The virtual bumper exploits the fact that the motors are not that powerful. If the maximum power (100% duty cycle of PWM) is applied to the motors for 1.5 second and robot is not moving then collision is registered. Again, this detection has proven to be quite reliable given the combination of the robot weight and motor power. From the view of the higher control software collision detected by the virtual bumper is indistinguishable from regular collision triggered by a micro-switch.

Obstacle Avoidance and Navigation Strategy

The obstacle avoidance is very simple, because a random walk (our coverage principle) does not require precise realization of the plan. When collision occurrs robot tries to roll-back its last action. If it collides while moving forward, it backs up 25cm and turns left or right for given value. The values are precomputed and stored in a table. If it collids while turning, it returns to its original position (the robot always turns on a spot), and backs up another 25cm. The last case, when the robot collids while reversing is not properly handled. It tries to turn, but better solution would be (based on experience in Lausanne) move forward say 10cm, and try to turn some smaller “escape” angle.

Coverage Principle

The presented strategy during the cleaning contest was a "random" walk. The rules were created by random generator, but for actual run they were coded into a table. Each command was defined distance how far should robot move straight and how much it should turn (with a sign determining direction left/right). This strategy was chosen for its simplicity and robustness (taken into account the environment and sensors used).

Performance Data

The performance was measured separately for cleaning, obstacle avoidance and for coverage:
The test track for cleaning was on linoleum floor covered with confetti and sugar. The nozzles and the brush were then adjusted according to the immediate results.
The obstacle avoidance was tested in normal office environment - chairs, tables, cables on the floor. It showed that simple bar with two micro-switches as a sensitive bumper is quite sufficient for given task. The maximum speed and position of the bumper was tuned during these experiments.
The coverage was tested only under simulation. The main control loop was called 600000 times (10 minutes, 1ms cycle), and coverage ratio was based on the number of squares visited by robot. This was not perfect evaluation, but it was a good indicator if the strategy is getting better or worse.
People
Martin Dlouhý (md)team leader, SW, testing
Jan Hradil (jh)HW, brushes, pizza box
Mike Hibbett (mike)vacuuming box
Ludvík Tesař (lt)HW, brushes, pizza box
Jana Halgašová (jana)surface finish, presentation