• Tidak ada hasil yang ditemukan

ADAPTIBILITY AND LEARNING

EXERCISES

4.3 ADAPTIBILITY AND LEARNING

For a human being, repeating a job is usually easier than doing it for the first time and such repetitions helps to develop special faculties to execute the job without any real hassles. This enhances performance and therefore learning is an essential tenet of AI agency.

For autonomous agency, Toda’s fungus eater model, Wilson’s model of ANIMATs and Pfeifer’s design paradigm forcefully make their contention towards adaptability for long term performance over an assortment of tasks for the robot. Therefore, robots must be designed to learn from its experience of interacting with the environment.

As an ab-initio, learning can be seen as a mapping from a sensed world (sensors) to the

chapter, hybrid architectures have been the choice of the designers to overcome such variety and dynamism offered by the real world. As an alternative choice to hybrid architecture, incremental learning has also found favour. This is because, (1) learning is easier to implement and nowadays in most of the robotic software suites it is provided as a ready made module in the robot’s firmware, (2) unlike behaviour which is pre-programed, learning happens on the run and lends itself as a control policy and is more atomic in execution than behaviour and therefore (3) doesn’t need to be designed separately as a behaviour nor does it need an executive for coordinating various behaviours, (4) learning works out across simple behaviours to the more abstract ones - even those which the architecture may not have accounted for and one doesn’t need to resort to semi empirical mathematical models of the environment or time varying functions of the uncertainty which may be encountered by the robot.

From an engineering point of view, a robot given a specific task and more or less contains itself to a known environment, the robot designer knows the a priori information and representations needed by the robot, and unless there has been severe and sudden changes to the environment or the robot, in principle will not need to resort to incremental machine learning. However, researchers have also put attention to particular representations, as saving battery power, avoiding falling off cliffs and avoiding threatening natural formations, covering a certain stretch of terrain or finding a specific item as mineral ore, vegetation or a known landmark etc, so that repeating these feats become easier in the apparent future.

Learning can be manifestly employed in a robot to either provide apriori representations and information which will enable it either to readily interact with the world in a certain manner than to learn it afresh over time over a great number of attempts, or to enable it towards specific goals and help it survive for longer durations in difficult terrain. For simple behaviours as navigation, learning will help to modify representations of the world which will simplify the given task. Since low level reactive tasks are closely related to navigation, hence making a map becomes important for tasks which need repetition. For example, a sub goals technique which leads to the final goal will need the robot to learn to recognise the subgoal, viz. if the subgoal is marked with a blinking red light, the robot will need to behave nearly the same to all such subgoals, until the final goal is reached. Learning is closely related to intelligent behaviour and autonomy, and not only helps the agent to acquaint with the environment, but also is a remedy for sensor noise and faulty and damaged sensors and actuators. In various robot architectures, learning works in tandem with the control loop.

Learning for sophisticated tasks over long term implementation will lead to development of newer and advanced behavioural modules. As an example, the working of guard dog robot in an unmanned facility is briefly shown in Figure 4.17, here the robot is provided with simple behaviours and it has to adhere to a known calibration, that of the circuital path

151

FIGURE 4.17 Learning for a guard dog robot, considering a robot patrolling an unmanned facility on a far off planet or at a difficult terrain which is difficult to access as a mountainous area or a hazard prone area. Here, the robot has go around the given circuital path (around the big grey glob) and sends photographs and information back to the the mission command at the dotted points (subgoal point) on the map. The loop taken by the robot (Flag point-1 to Flag point-10) for surveillance is nearly the same for all its pursuits and machine learning techniques (as ANN, Fuzzy logic) helps to keep the robot to its path for each loop and helps to avoid changes due to the diminishing onboard power, uneven terrain, weather conditions, environmental factors as storms, wear & tear, lack of regular maintenance etc. The robot was designed for simple behaviours as,

‘motion to (sub)goal’, ‘obstacle avoidance’ and ‘take photograph’, long term surveillance is a new and superior behaviour developed by machine learning techniques.

by using machine learning techniques, and over time it develops surveillance as a superior behaviour.

To summarise, learning helps to acquire the following features to robots [54]:

1. Adds in utility of particular representations to enhance performance, for the given example these special representations are the dotted points which serve as subgoals.

2. Helps relate special instances of sensor for action, or behaviour coordination, for the guard dog robot this sort of calibration between sensor reading to action can be seen, that it has to take photographs on reaching the subgoal.

3. Overcoming uncertainty, this is more appreciable if the robot is engaging in repetitive tasks in a predictable domain. As discussed, the machine learning techniques ties the circuital path to the robot’s performance.

4. Gradual development towards newer and more involved behavioural modes, for the given example this is the development of surveillance as a new behaviour by melding simple behaviours.

The first application of ANN for robots led to improvements to the design of earliest models of autonomous vehicle. The NAVLAB vehicle was the outcome of research into autonomous vehicles in the late 1980s. Autonomous Land Vehicle In Neural Network (ALVINN) was the brainchild of Dean Pomerleau [268]. ALVINN was put to work on the NAVLAB vehicle. It was a refurbished army ambulance fitted with a 5,000 watt generator and had an operating system which could deal with 100 million floating point-operations per second. ALVINN had a three layered back propagation neural network which took images from a camera and data from laser range finder as input and computed the direction the vehicle should travel to continue to follow the road as output. Training was done using

FIGURE 4.18 ALVINN architecture, was a three layered back propagation neural network the images from a camera and data from laser range finder were the two inputs and the intensity was used as a feedback, it computed the direction the vehicle should travel to continue to follow the road. Image adapted from Pomerleau [268].

simulated road images. Since the vehicle did not follow a predetermined path and the local variables guided towards the direction the vehicle should follow, it was adaptable across a variety of conditions and over various types of terrains. The highest speeds attained using ALVINN on the NAVLAB vehicle was about 70 mph.

Evolutionary approaches for developing control systems for robots can be done either by using real robots or by using software. However, using real robots it is seen that the even the simplest behaviour models tends to run over a hundreds of generations spread across a very large number of hours, therefore power is therefore the most obvious issue, as batteries will not sustain for such long and tethering using a power source may not be a suitable option at every situation and may limit the process. Jakobi also points to wear and tear of the robot, which can be substantial and real robots is not a practically viable option. Hence, the preferred route is via software, however with sophisticated robots and more demanding behaviour the semblance between the simulation and the real robot’s performance is hardly any. Therefore, the onus was to develop a methodology on which simulators can be easily and cheaply built that enable that both the simulations and real robots to affect the nearly the same performance, which is scalable in time, speed, number of robots and complication of behaviour domains. In his revolutionary approach, Jakobi suggests development of base set of robot-environment interactions and also a base set aspects of the simulation. Fuzzy logic has been an easy to implement machine learning technique for robots. Using fuzzy logic in tandem with a navigational algorithm as potential field has been a favourite technique for researchers.

Another technique is to implement case based methods. Most software driven systems are backed up by a memory cache. Designing behaviour using cache memory leads the agency to respond to the same situation in nearly the same manner as it did previously. Case based

153 approaches [271, 272] propose to make an exhaustive list of scenarios encountered by the agent and its response and update this list if and when newer situations are encountered.

Case based learning can be summarised in the following five steps;

1. Define the problem at hand

2. Obtain similar case(s) from memory

3. Modify these case(s) to suit the problem at hand 4. Apply the solution developed, and evaluate the results 5. Learn by storing newer cases

This approach works very well for navigation and low level behaviours typically for a simple world and is a proven improvement over pure reactive methods.

SUMMARY

1. Navigation and adaptive learning are two important tools for a roboticist.

2. Navigation is the simplest behaviour and the easiest test bed to design newer and richer behaviours.

3. Grid based methods are easy to implement but they lack in reactivity and cannot be implemented in a dynamic environment.

4. Potential fields method has been particularly popular for autonomous navigation.

5. ND is a navigation algorithm based on situated activity and is very successful in 2D navigation.

6. Simple navigation algorithms cannot take into account the shape of the robot and the kinematic considerations.

7. There is a lack of good navigation methods in three dimension.

8. Adaptive techniques brings forth newer behaviour in the robot by robustly accomodating for the changes in environment.

NOTES

1. Potential fields method finds applications in the gaming industry.

2. A variant of VFH has been used to make ‘Guide Cane’, a robotic navigation guide for the blind.

3. A? is not recommended for sensor based navigation, as it depends on heuristic evaluation of a given node of a grid, and may lead to discontinuities as the robot traverses from one node to another.

4. Braitenberg vehicle-5 is motivated from the Turing Machine and is said to have a

‘brain’, this is achieved with threshold devices plugged between sensors and motors.

Some threshold units will fire only when the signal is above threshold, while others will only stop firing when threshold is exceeded, in effect establishing two states of operation viz. ON state (ONE) and OFF state (ZERO) tailored by the threshold.

Blending excitatory and inhibitory connections from vehicle-3 for the threshold units, Braitenberg very nearly replicates the McCulloch-Pitts neurons where the threshold devices are sames as hidden neurons, and shows that when arranged into a feedback loop these devices can be used to instantiate simple memory units, akin to digital logic (AND, OR and NOT). However, the computational capacity of the vehicle is rather limited, and 10 threshold devices it can store only a 10 digit binary number or a maximum decimal count of 1023. Vehicle-5 is designed not only to realise the

implemented to charter unknown and dynamic environments.

EXERCISES

1. Grid based methods, demonstrate with examples that grid based methods leads to suboptimal paths.

2. C-type obstacle for ND, discuss the performance of ND diagram navigation for C-type obstalces.

3. Local minima, show that local minima problem for potential field navigation can also occur when the obstacle is not C-type or U-type. Based on this results, is potential field navigation a good choice for generic navigation with uneven terrain, moving obstacles and other robots?

C H A P T E R

5

Software, simulation &

control

“ROS = plumbing + tools + capabilities + ecosystem”

– Brian Gerkey

“If you could see the world through a robot’s eyes, it would look not like a movie picture decorated with crosshairs but something like this:

225 221 216 219 219 214 207 218 219 220 207 155 136 135 213 206 213 223 208 217 223 221 223 216 195 156 141 130 206 217 210 216 224 223 228 230 234 216 207 157 136 132 . . . ”

– Steve Pinker