ASSEMBLY
ROBOTICS
GROUP
Part of the
Institute of Perception, Action, and Behaviour,
in the
School of Informatics,
University of Edinburgh
Sister Groups:
computer graphics
machine vision
mobile robotics
For more information contact: Chris Malcolm
Last updated: Mon May 1 14:47:17 2000
|
|
The Assembly Robotics group in Edinburgh has a long and distinguished
history. Some of the highlights are mentioned below:
Currently the focus of our research concerns is finding more
principled support for the familiar intuitive general guidelines of
behaviour-based robotics, and exploring its use in the hybrid version
of behavior-based robotics we use in assembly work (a hybridisation
of classical planner and behaviour-based plan interpreter), with
special emphasis on the sensor fusion problem.
Freddy, the Famous Scottish Robot
Freddy (mid1960s - 1981) was one of the first robots to be able to
assemble wooden models using vision to identify and locate the parts
-- given a jumbled heap of toy wooden car and boat pieces it could
assemble both in about 16 hours using a parallel gripper and single
camera (1973). The 16 hours was due to the slowness of movement of the
robot, an artefact of the limited computational power available for
movement control in those days. An Elliot 4130 computer with 64k
24-bit words, later upgraded to 128k, was the main computer. A
Honeywell H316, initially with 4k 16-bit words, later upgraded to 8k,
controlled the robot motors and cameras. The videos we now have of
Freddy's assembly work have been dubbed from 16mm film, as video
hadn't been invented then. Even with today's knowledge, methodology,
software tools, and so on, getting a robot to do this kind of thing
would be a fairly complex and ambitious project. In those days, when
they had to design and buld the robot, design and build the
programming system, design and build the vision system, etc., it was a
heroic pioneering feat which required to be demonstrated in practice
in order to convince some that it was even possible.
Key Reference
A. P. Ambler, H. G. Barrow, C. M. Brown, R. M. Burstall, and R. J.
Popplestone, A Versatile Computer-Controlled Assembly System,
Proc. Third Int. Joint Conf. on AI, Stanford, California, pp. 298-307, 1973.
Freddy's famous car and boat asssembly above was programmed as a list
of end-effector positions. The great tedium and lack of generality in
programming assembly robots in this way prompted the search for a
higher level of assembly description. RAPT permitted robot positions
and movements to be specified in terms of relationships (such as
parallel, aligned, against) between geometric features (such as point,
edge, and surface) of the parts being assembled. By 1980 this was well
developed (largely by Popplestone, Ambler, and Bellos) and had been
integrated with a solid geometric modeller front end (by Cameron) to
facilitate part description and trajectory planning. The geometric
nature of vision permitted its neat integration within the RAPT system
using such ideas as a plane-of-gaze (defined by lens centre and two
points in image plane marking an object edge) projected out and
touching the edge of the real object (by Yin). The attempt in the mid
1980s to include reasoning about uncertainty based on tolerances on
dimensions and positions, errors in robot movements, etc., foundered
on a combinatorial explosion of computation.
This raised the interesting question of whether this was an
unavoidably hard problem which simply needed very much more powerful
computers, or whether there was another computationally cheaper way of
tackling the problem.
Key Reference
Popplestone, R.J., Ambler, A.P., and Bellos, I., An Interpreter for a
Language for Describing Assemblies, Artificial
Intelligence Vol. 14, No. 1, pp. 79-107, 1980.
|
The SOMASS system was intended to overcome the computational problems
of dealing with uncertainty at the geometric modelling and reasoning
level by dealing with it at a lower level in the system, by means of a
behaviour-based assembly plan interpreter, the plan having been
constructed by a classical assembly planner which imagined it was
dealing in an ideal world with no uncertainty. This changed the
difficult problem of working out all possibilities in advance in the
off-line system, to the easier problem of handling the small subset of
these possibilities actually encountered by the on-line system when
running. An idealised and certain world was presented to the planner
by the capability of the behaviour-based plan interpreter to deal with
expected amounts and kinds of uncertainty. Planners are good at
handling perfect idealised worlds. Instead of trying to develop the
planner to handle an imperfect realistic world, which had proved
computationally intractable when attempted as an extension to RAPT,
attention was shifted to finding a way of using perfect idealised
plans to handle imperfect realistic situations by developing the
capabilities of the plan interpreter.
The behaviour-based nature of the plan intepreter and robot controller
was akin in general philosophy to the behaviour-based ideas being
developed by Brooks of MIT in the domain of simple reactive mobile
robots at the same time, but here were designed to act as an assembly
plan interpreter. In the early 1980s Brooks (with Lozano-Perez) had
been working on the general strategy for developing an ambitious
assembly system incorporating planning and uncertainty
handling. Coming to realise that this strategy was doomed by its
combinatorially explosive computational requirements he decided that
robotics research should switch to small simple reactive but complete
mobile robot systems. These would be incrementally developed into
robot systems of greater complexity and scope in an abbreviated
recapitulation of the progress of biological evolution from simple to
more complex animals. This coupled animal biology and ethology with
robotics in what is sometimes called "biologically inspired robotics",
a programme of research which saw robots as synthetic analogues of
animals, picking up a tradition pioneered by Grey Walter in the 1950s.
The collapse of the ambitious assembly robotics systems with
integrated planners was seen as demonstrating clearly that robotics
researchers had got things very badly wrong. The idea was that
starting with very simple mobile robots "animals" and working upwards
like this would enable robotics researchers to learn the fundamentals
of robotics, this time correctly. The rejection of what was now
(mid-to-late 1980s) seen as the failed "classical" robotics
architecture, based on "classical" or knowledge-based AI led to a
general horror amongst the new breed of roboticists of knowledge-based
classical AI. Since plans and planners were paradignatic examples of
classical knowledge-based AI, and assembly robots had to have detailed
plans of how to put things together, assembly robotics was generally
seen as an unprofitable area in which to pursue robotics research.
The exception to this was Edinburgh, who saw behaviour-based robotics
of this new kind as a way of revising the division of labour between
planner and plan-performing agent in such a way that their marriage
become a profitable and computationally economical collaboration,
instead of the combinatiorial explosion into intractability of the
previous "classical" approach. The idea was to marry a classical
idealised planner with a behaviour-based plan interpreter.
The chosen experimental domain to test this hybrid architecture was
the SOMASS puzzle, a kind of simple 3D jigsaw puzzle based on seven
bent bricks which could be assembled into a cube. This had been
invented by the mathematician Piet Hin, allegedly during a boring
lecture by Heisenber. He called it the "SOMA puzzle". It became a
popular puzzle and is available at all good puzzle shops. The first
version of the SOMASS (SOMa ASSembly) system (1985) used no sensors,
dealing with uncertainty by using compliant motions, and was capable
of planning and performing the assembly of a SOMA cube in dozens of of
different ways. This was generalised to handle the assembly of any
shape from these parts, and from instances of these parts of any size,
and by 1990 behaviour-based uncalibrated vision-guided part aquisition
had been developed, which did not need to know camera parameters or
position, or robot kinematics. What a system does not need to know it
does not need to worry about, and things the system need not know are
free to change without affecting the operation of the system. By way
of humorous contrast with knowledge-based systems, this was sometimes
jokingly referred to as an ignorance-based system.
Key References
Chris Malcolm,
A Hybrid Behavioural/Knowledge-Based Approach to Robotic
Assembly, Evolutionary Robotics '97, April 17-18, Tokyo,
Japan, ed Takashi Gomi, 221-256, AAI Books, Ontario, Canada,
1997. Edinburgh University DAI RP 875.
Back to top
``What if all these problems of uncertainty which so bedevil assembly
robotics were an artefact of trying to do assembly with a
position-controlled robot?'' This was the question which Graham Deacon
asked in 1994. To answer this question he built a direct drive (and
therefore easily back-drivable) robot arm which was purely force
controlled. The hope was that once parts to be assembled together had
been brought into initial contact, it would prove possible to devise
part fitting routines based purely on force control strategies,
completely ignorant of position, which would accomplish typical
assembly problems such as peg-into-hole and brick-into-corner. If this
proved possible, then a peg-in-hole routine (for example) which
successfully put a small peg in a small hole in one position, should
equally well, with no change, put a large peg into a large hole
anywhere else. In other words, the kind of family generality of
assembly task which others had sought with such high level
descriptions of the task as RAPT would in this case fall directly out
of the basic bottom level of control of the robot. If it knew nothing
about position, and needed to know nothing, it was not going to be
worried by positional errors. Another example of the ocassional virtue
of ignorance.
This proved to be the case, but programming the robot in terms of
force control routines was a difficult and expert task. Could any way
be found to simplify the programming of the force control routines?
While people find it particularly difficult to describe force control
tasks in formal terms, they are particularly good at learning the
knack of accomplishing a new force control task. Would it be possible
to equip a person with appropriate sensory feedback so they could
exploit their own natural skills and intuitions in learning how to
control Eddie (the compliant robot) to accomplish a certain task, and
would it be possible for the computer system to learn from observing
this process how to accomplish the task autonomously, unaided by human
control?
This is an ongoing research project, which has now shifted (with
Graham Deacon) to the Mechatronic Systems and Robotics Research Group
of the University of Surrey. The results so far, in the few tasks for
which this has been tried, are encouraging.
Key References
Deacon,G; Malcolm,C A,
Robot System Designed for Task-Level Assembly, Proc of
"International Workshop on Advanced Robotics and Intelligent
Machines", Salford, England, March 1994.
Deacon, G,
Accomplishing Task-Invariant Assembly Strategies by Means of an
Inherently Accommodating Robot Arm, PhD Thesis, 1997.
Back to top
Pushing and Sliding
A key element in the implementation of robot control is goal-seeking,
which can be implemented in many ways. Watt's famous steam engine
speed governor, which used centrifugal force to adjust the steam
valve, was a purely mechanical implementation. In the 1960s we would
have been more likely to solve this problem by electronically sensing
the engine shaft speed, and using a differential amplifier to subtract
this from the required speed, amplifying the difference and applying it
to (say) a sprung solenoid controlling the steam valve. In the 1980s we
would have been more likely to solve the problem by shifting the data
into a computer, evaluating a control function of whatever complexity
we cared to program, and using the result to control the steam
valve. Thus this problem can be solved mechanically, electronically,
or computationally.
Since robots are built from components from mechanical, electronic,
and computational realms, they can take advantage, when possible, of
the opportunity to solve this kind of ``computational'' problem
outside the computer. Not all problems can be solved at all these
levels of course; that's why digital computers ousted analogue
computers. But when they can, the result is often both a faster
response, and a lessened requirement for computational speed.
In assembly robots, where it is often possible to engineer the parts
and the environment to suit the task, one can often save computation
in this way by taking advantage of local physics, such as by
exploiting guiding chamfers on the edges of pegs and holes, or using
gravity's handy tendency to align the bottoms of objects with the tops
of tables. In general, a great deal more than many imagine can be
done by partly constrained and partly compliant motions of robot and
parts, such as in pushing and sliding, and we (largely Deacon and
Wright) have contributed to the theory of pushing and sliding.
Key References
Deacon,G; Low,PL; Malcolm,CA,
Orienting Objects in a Minimum Number of Robot Sweeping
Motions, Edinburgh University DAI PP 686.
Deacon,G; Wright,M; Malcolm,CA,
Qualitative Transitions in Object Reorienting Behaviour, Part 1:
the Effects of Varying Friction, Proceedings of IEEE
International Conference on Robotics and Automation, pp.2697-2704,
Albuquerque, New Mexico, USA, April 1997, Edinburgh University DAI RP
863.
Deacon,G; Wright,M; Malcolm,CA,
Qualitative Transitions in Object Reorienting Behaviour, Part 2:
the Effects of Varying the Centre of Mass,, as
above, Edinburgh University DAI RP 864.
Back to top