Home   |   Members only   |   Contact   

Site menu:

Latest news

2013.08.31
Official end of project

Press release on concluding public workshop

See documents »

Presentations and photo's from concluding workshop

Public workshop »

Updated movie of autonomous tractor

See movies »

Location: Lecopro.org / Results 

Results

On this page you can find an overview of the development cases studied in LeCoPro and a selection of the project results achieved on these cases.

Development cases:


Transmission

Badminton demonstrator

Tractor with implement

Tractor in LEGO MINDSTORMS

Sliding mass with optical sensor

Hydrostatic drivetrain

Shooting ball setup-up

Transmission

In this development case, learning control has been applied to the engagement of wet clutches, which are components typically used in power transmissions of off-road vehicles. One difficulty for control is that these clutches have a hard non-linearity as no torque is transfered until a hydraulic piston makes contact with a set of friction plates and presses them together. Another difficulty is that such clutches are used in strongly varying environmental conditions. Both model-based and model-free methods have been developed to address these issues, each of them learning the control inputs that achieve an optimal tradeoff between an engagement that is fast on the one hand, and smooth and comfortable on the other. The developed techniques have been validated on a dedicated test bench (Fig. 1), where an electromotor (30kW) drives a flywheel (25 kgm2) via a torque convertor and two mechanical transmissions.


Fig. 1: Transmission test set-up


Selected results:

Top

Badminton demonstrator

To demonstrate its mechatronic competences, FMTC has constructed a badminton robot (Fig. 2). A visual system with two high-definition black and white cameras is used to localize the fast moving shuttle. The robot itself consists of one linear motor and two rotational motors. In the LeCoPro project, a variety of learning techniques has been applied to the motion control of these different motors, both for normal motion control and control of repetitive serve operations.

A movie of the badminton robot in action can be found here.


Fig. 2: Badminton demonstrator


Selected results:

Top

Tractor with implement

This development case consists of a small tractor (Fig. 3), that can be controlled automatically (steering, speed), and a carried implement, that has rear wheels to steer relative to the tractor. In a first phase, learning control has been applied at the subsystem level to control the tractor, while in a second phase the entire system consisting of two subsystems (tractor + implement) has been considered. The main challenges in this application are the varying soil conditions and the interaction between the tractor and the implement. Both issues have been addressed by the addition of recursive learning laws that allow to iteratively build up a model for these effects, which is then used by a model-based controller to realize the desired accurate motions. Besides this, also several model-free techniques have been developed which automatically learn an optimal controller through interaction with the controlled system.

A movie of the tractor and implement action can be found here.


Fig. 3: Tractor with implement


Selected results:

Top

Automous tractor built in LEGO MINDSTORMS

For the ease of developing model-free controllers in a safe environment, a scaled version of the autonomous tractor has been built in LEGO MINDSTORMS. The result is shown in Fig. 4, and a movie of the tractor in action can be seen here

Autonomous tractor built in LEGO MINDSTORMS
Fig. 4: Scaled version of tractor built in LEGO MINDSTORMS

Top

Sliding mass with optical sensor

A fourth development case deals with the motion control of a sliding mass which is excited by the impact of a mass-spring-damper system (Fig. 5). Only a single optical sensor is installed to check when the excited mass passes a certain position. In a first task, the input signal to the motor actuating the mass-spring-damper system had to be controlled such that the impacted mass follows a desired motion, for example passing by the sensor at a given time instant. Learning techniques have been developed to realize these motions, in order to demonstrate the potential of learning for systems where the controlled output can not be measured continuously, but only at discrete positions. In a second task, identification techniques have been developed which use the binary data as measured by the optical sensor, and from this estimate a model able to predict the continuous motor position based on the motor input signal.


Fig. 5: Sliding mass with optical sensor


Selected results:

Top

Hydrostatic drivetrain

A fifth development case is a hydrostatic drivetrain (Fig. 6), consisting of an electromotor, a hydraulic pump and two hydraulic motors. Heavy duty mobile machines such as road construction vehicles, agricultural and forestry machines are frequently driven by this type of drivetrain. As this development case consists of several strongly interacting subsystems both in the electrical, mechanical and hydraulic domain, it is and was an ideal development case for global learning control algorithms.


Fig. 6: Hydrostatic drivetrain

Top

Shooting ball set-up

A sixth development case is a set-up consisting of two linear motors (Fig. 7). The first one is located 0.5m above the ground. It pushes a ball from standstill to a particular speed over a rail to launch it in the air. The second one on the ground is connected to a basket to catch the ball and bring it back to the position below where it is launched. The purpose is to launch the ball, catch it and bring it back to below its launching position, in such a way that the total time to perform this movement is minimal. A trade-off has to be made between the durations of the first motor movement and of the second motor movement. Indeed, the first movement is shorter if the ball is launched at a higher speed. But then the ball has to be caught at a longer distance, increasing the duration of the second motor movement. The overall optimum depends on the maximum reachable speeds of the two motors. The total movement can either be learned in a model-based way or in a model-free way. The model-based solution consists of iteratively updating friction estimates and adjusting the launching speed and catching position. In the model-free way, two independent reinforcement learning agents learn the launching speed of the first motor and the catching position for the second motor.


Fig. 7: Shooting ball set-up


Top