Knowledge Based Error Recovery in Industrial Robots

1. Pick up component

2. Transfer component

(coarse, fast movement between locations)

3. Place component

approach

insert

ungrasp

depart

(fine movement towards target

location)

(fine movement while in physical

contact)

(release component)

(fine movement away from target

location)

The "locate action" is not a robot function but refers to the action of either part feeders (implicit location data) or sensors (explicit location data).

Now by examining each of these atomic actions in turn we obtain a tree of error states and their relationships. For example:-

Pick up

Location error

Approach error

Grasp error

Depart error feeder empty or jammed part defective part orientation error collision with part missed part collision with workspace no part several parts orientation error

• part location error
• gripper slip
• obstruction
• part error
• grasp error
• workspace error

This is only a sample of the full tree: there are further levels and more detailed expansion. By using the concept of an "atomic action" we have produced a detailed analysis for each manipulator movement axis [5]. These trees illustrate the inherent structure in the error states and their causes. If we have knowledge of the current action and sensory data then, from the tree, we can infer plausible error causes. Several iterations may be required, involving requests for selective sensory data.

V KNOWLEDGE BASE DESIGN

The ideal knowledge base for a robot would contain geometric data in a body modelling system which could represent the exact physical structure of the workspace and the objects. By using computational geometry and extensive high powered sensing the system would deal with contingencies by generating its own recovery plans [6]. However, while research is active on these topics, such object level systems are not yet well developed and will be very complex and expensive.

Our aim is more modest: to develop a knowledge base of actions and events in terms of the task descriptions used in existing manipulator languages. This is more a task event model than a body modelling system. The original task specification consists only of a sequence of VAL instructions, our final specification will have several parallel levels of ancillary knowledge cross-referenced into the VAL program. This multi-layered approach is a key feature of successful knowledge bases [7]. Our current levels are as follows:

• Task level - VAL program (manipulation cycle of actions and tests).
• World level - Knowledge of objects in environment.
• Action level - preconditions, actions, effects and results.
• Diagnostic level - sensory interrogation procedures.

Recovery level - parameterised recovery schemas.

Other levels being examined include a user level, performance level and a meta strategy level.

Our current activities are directed at:­ knowledge representation (frames and/or production rules), properties of objects and events (e.g. reversability), and the separation of generic diagnosis and fault data from application specific data. The next stage will be to implement and test our design in the application framework.

ACKNOWLEDGEMENTS

We are grateful for the support of British Robotic Systems Ltd. and the S.E.R.C. through Grant GR/B/94106.