Architectures for Controlling an Intelligent Autonomous Vehicle

Download 1.23 Mb.

Page	6/7
Date	03.05.2017
Size	1.23 Mb.
	#17133

1 2 3 4 5 6 7

Figure 5.5

51

Five features of the performance of this combined architecture deserve further note. Three of these features point up the strengths of this combined architecture. First, while using this architecture, Ford successfully achieved goals in 70 of the test situations. When confronted by an obstacle that could be avoided or circumnavigated, Ford used a combination of obstacle avoidance and wall following techniques to return to the path for its goal (see Example 1 in Appendix L). When a temporary diversion proved impossible and the map provided a second route to the goal, re-planning occurred to find that route (see Example 2 in Appendix L). Second, the ability of the system to plan and re-plan to achieve a goal gives the this architecture a far greater level of flexibility in achieving goals than the subsumption architecture. The additional features enable the vehicle to locate goals which are outside its line of sight. Third, at least in these experimental conditions, the topological map appeared to effectively represent the environment using the position of landmarks without a need for resorting to dead reckoning.

Nevertheless, two features of the experimental results highlight deficiencies with the present construction of the combined architecture which will require adjustment in future work. First, as a consequence of the inadequate level of sensory input fed to the processor and the simplicity of current sensory processing techniques, the vehicle sometimes got stuck in oscillating behavioural loops when it approached near objects, causing an unnecessary request for re-planning. The other major effect of the sensory deficiencies is that the vehicle lacks the ability to detect dead ends and to initiate re-planning prior to experiencing entrapment. Second, this combined architecture retains a deficiency of the subsumption architecture tested in Chapter Four. The subsumption element prevented the reordering of behaviours to enable the IA V to touch its goals, while still avoiding obstacles en route.

In summary, the combined architecture exhibited a high degree of success in enabling Ford to achieve goals. The next section details proposed modifications to the implementation of this architecture which may enable at least partial rectification of the two remaining deficiencies.

52

5.4 Future Work

Based on the work completed in this chapter, I see three major areas for future development. First, the results of the experiments with the combined architecture reinforce the conclusion in the previous chapter that further testing of subsumption should be conducted to determine if this architecture can be modified to reorder the network of behaviours to achieve differing goals, such as avoiding obstacles but touching beacons. Second, both the sensory devices on the base vehicle and the sensory interpretation software require improvement. Again, as with the results of the previous chapter, these experiments further support the need for the use of visual input and vision processing, along with other more reliable and sophisticated sensory techniques, which would enable the vehicle to recognise landmarks and create its own topological maps without the aid of pre-programmed maps. Hopefully, this work may also yield a means for IAVs to more effectively cope with dead ends.

Third, one can experiment with the use a second plan-then-execute architecture as a still higher level in the hierarchy to produce lists of goals to achieve tasks such as delivering mail. Figure 5.6 provides a diagram of such an extended architecture. This extra plan-then-execute architecture would use a non-linear planner, which would give this higher level of the hierarchy the power and flexibility to achieve a wide range of goals. I have already implemented the non-linear planner in PROLOG, but have yet to upgrade the rest of the architecture with the above-specified modifications. Many, if not all, of the above detailed improvements will be completed in the next year as part of a collaborative project with British Telecom.

5.5. Summary

In summary, the experiments detailed in this chapter demonstrate the compatibility of subsumption and a scaled-down version of a plan-then-execute architecture. This research holds out the possibility that a combined reaction-based architecture and a fully implemented plan-then-execute architecture may yield yet more impressive results. Likewise, these experiments suggest that a combination of the

53

: (Infra Red Beacon Number) I 1 _ . I : Execution ~ -loE--- I Controller : -----r---- V : Plan then Execute Module 1 ____ J_____ I ~ 1 1_M, : Topological ~ ~ appmg :.:_ ...•. __ --- : Planner I : Module : -----r----^I ---------- : (Infra Red Beacon Number) 'iJ 1 1 1 _ • 1 : Execution ~ ...•. __ --- 1 Controller : -----r---- V Plan then Execute Module 2 1---------- I-----~---~ : Non-Linear ~ ~ Mappmg :.:_ ...•. __ --- : Planner: ~ _ ~.?~~l~ : -----r---- Stepper Motors						I:J
				Find Infra-Red				Jnfr

						Module		Se

			Obstacle					UltraSound

			Avoidance					Sensors

	~	Reflex Module					_Bumper
	~				__		_Bumper
~							Bars

Figure 5.6

aRed nsor

54

All Sensors

All Sensors

All Sensors

All Sensors

₁

programming languages 'e' and PROLOG can produce an efficient implementation of a control system. This combined architecture did enable Ford to find beacons, though these experiments again highlighted the need for improved sensory capacity in the Marvin vehicles. The future work which is now being funded by British Telecom may likely facilitate the improvement of both the IA V hardware and this combined architecture to allow the IA V to perform a more complex series of tasks.

55

Chapter 6

Conclusions

Contents

6.0 Introduction

6.1 Conclusions About Hardware

6.2 Conclusions About the Implementation of the Architectures 6.3 Closing Notes

6.0 Introduction

In comparison with the constructed intelligent beings of Gene Roddenberry's Star Trek, Isaac Asimov's Robots and Empire series, or other technologically-oriented science fiction, the intelligent mobile vehicles of today's laboratories are mere rudimentary toys. Nevertheless, humans have made a remarkable achievement by

creating vehicles which can exercise basic levels of judgement. Biological beings no

longer hold an exclusive monopoly in the field of decision-making. The present research into robotic architectures, mapping systems, and planners - the modules which empower an IA V to perform basic but essential levels of reasoning now exercised by (relatively) simple biological species - is enabling computer scientists to compile a foundation of knowledge which scientists in the future may be able to use in the

56

creation of beings capable of independently contributing to the theoretical foundations of technological research.

Rather than addressing these lofty speculative issues, this dissertation has contributed to the practical discussion of the implementation of programs enabling IA V s to perform simple behaviours by fIrst reviewing the debates in the ArtifIcial Intelligence literature concerning IA V architectures, mapping systems, planners, and attention; and then by reporting the results of tests of three different architectures: a reaction-based-architecture; subsumption; and a combined subsumption/plan-thenexecute architecture using a topological non-linear path planner. While the limitations of time and resources associated with MSc research necessarily restrict the scope of conclusions which may be drawn from this dissertation, the results of this research does suggest some modest conclusions about IA V hardware and software.

6.1 Conclusions About Hardware

Two different IA V s served as base vehicles for this project, a small and simple vehicle designed and produced in the course of this research, and Ford, a Marvin vehicle developed in the Department of Computer Science at the University of Essex. Observation of the operation of these two vehicles support three conclusions. The simple vehicle described in Chapter Three required inexpensive hardware easily purchased on a student budget for construction. This vehicle successfully roamed the research laboratory. The research detailed in Chapter Three thus fIrst suggests that vehicles capable of performing simple tasks may be readily designed and produced at a minimal cost. I assigned a series of more complex tasks to Ford, and found that the ultrasound and infra-red sensors with which this vehicle is equipped did not provide suffIcient information about the environment to enable this IA V to interpret its surroundings under some conditions. The addition of further ultrasound sensors improved the vehicle's performance, but did not substantially increase Ford's ability to interpret its environment. The research in both Chapter Four and Chapter Five suggests the second conclusion that the Marvin vehicles would operate more efficiently

57

if more comprehensive sensors, such as vision, complemented (if not replaced) the ultrasound and infra-red sensors. Third, the tests conducted for all three implementation chapters suggested that speed of movement is not necessarily a highly desirable initial feature for IA V s at the present time. The small vehicle presently navigates rooms at a speed slightly excessive for convenient observation. In the case of Ford, the maximum speed could not be permitted as the vehicle's sensors could not accurately feed information to the subsumption architecture at a rate to permit efficient movement at such a speed, and since the non-linear planner could not produce plans for goal achievement at a rate to permit the use of the higher potential speeds.

6.2 Conclusions About the Implementation of the Architectures

The implementation chapters review tests of a simple reaction-based architecture, then of subsumption, and finally of a combined architecture designed for this project. Each subsequent architecture facilitated more complex behaviours than its predecessor. The results of each chapter facilitate little direct comparison, as the base vehicle for the first implementation differed entirely from the vehicle working with the other two architectures, and as I implemented each architecture under different test conditions. For this reason, this section details the general conclusions about the implementation of these architectures individually.

Chapter Three shows how a simple mobile vehicle with a reactive control architecture can successfully perform a single task, such as obstacle avoidance. Nevertheless, the vehicle did not enjoy an unproblematic performance. The reactionbased architecture developed for this vehicle occasionally restricted this IA V to oscillating between states, though, as noted in this chapter, the use of an internal state in the architecture may possibly cure this problem. A second and more significant limitation of this architecture is the presence of only one layer of control. If the vehicle avoids obstacles, it cannot perform other behaviours, such as achieving a goal, at the same time. For this reason, I am one of many researchers who has experimented with other possible architectures.

58

---------------------------------- ------- ----------

The most sweeping claims made about any architecture appear in both the academic and popular literature discussing subsumption, an architecture developed by Rodney Brooks and his team at M.LT. For this reason, Chapter Four reports on the results of the retesting of the suitability of this architecture for achieving a simple goal, locating infra-red beacons. Generally, this architecture did produce successful results, however, the findings of the work conducted for this report do not support the entire range of abilities attributed to subsumption. Some difficulties likewise developed during the implementation of this architecture. First, some arrangements of dead ends ensnared Ford into oscillating between behaviours. Second, and in my opinion more importantly, I found that I could not achieve two different aims using this architecture. In future experiments, I plan to continue to test subsumption to see if a modification can be found which would introduce some flexibility into the hierarchy of behaviours required by this architecture. In spite of these two problems, on most occasions, the architecture enabled the integration of several different behaviours so that Ford could achieve an environmentally-based goal.

The structure of IA V architectures presently tend to fall into two distinct categories, as do the primary inputs on which the architectures rely and the types of actions that an IA V will likely perform. The fIrst category, reaction-based architectures, tend to use immediate sensory information to perform routine tasks. The second category, plan-based architectures, use predefined maps or maps generated during an IA V's exploration of its environment to plan strategies for achieving unfamiliar tasks. Chapter Five details the implementation of an architecture combining subsumption with a simplified plan-then-execute architecture using a non-linear planner which I designed to attempt to capitalise on the strengths of both categories of architectures and to attempt to mitigate the weaknesses of each. I noted the following five significant observations during the testing of this combined architecture:

59

The architecture successfully achieved goals in 70 of the test situations;
The ability of this system to plan and re-plan paths for task achievement permitted a higher level of flexibility than the subsumption architecture implemented on its own.
Ford successfully located landmarks in the environment using a topological map.
The limited level of sensory input received by this Marvin vehicle caused problems for the architecture.
The programming languages 'C' and PROLOG can be effectively interfaced to enable the architecture to respond to change, facilitated by 'C', and to plan and re-plan paths to goal achievement, facilitated by PROLOG.

6.3 Closing Notes

While human imaginations may well continue to outstrip our technological

capacity, computer scientists have nonetheless made remarkable achievements in the

study of artificial intelligence over the last few years. The research reported in this document generally supports the conclusion that the improvement of the flexibility of IA V architectures may well expand the capacity of thinking vehicles for task performance. This dissertation has also highlighted the existence of room for combining the features of the major trends in IA V research. While the evolution of thinking vehicles will not necessarily parallel biological evolution (Taylor 1994), the process of IA V evolution has undoubtedly only just begun.

60

Download 1.23 Mb.

Share with your friends:

1 2 3 4 5 6 7