Towards Automatic Error Recovery in Robot Programs

1. Example

At the beginning of the program the InitialModel can be :

Arm= ParkPosition HandOpening = X if TouchSensorTriggered then ObjectHeld, ObjectSize = HandOpening, ObjectPickedUpAt =Arm else ClearHand.

B. Semantics

We describe the semantics of the language in a STRIPS-ltke form (Fikes, 1971).

. Each instruction has associated a list of precondi­ ttons and postconditions. The preconditions express what should be true before executing the instruction, the postconditions express how to modify the current model after the execution of the action. They are expressed in term of additions (ADD) , deletions (DEL) , and updating (UPD) to the model.

1. Examples

We consider a small subset of AL instructions (Bin­ ford, 1979) .

MOVE ARM TO f r arne

prec :

post : UPD: amFframe;

OPEN HAND TO d

prec:

post: if ObjectHeld

then ADD : ClearHand

DEL: ObjectHeld,

ObjectSi ze =X . ObjectPi ckedUpAt UPD : Opening = d CLOSE HAND TO d prec : ClearHand post : UPD : HandOpening = d if TouchSensorTriggered then ADD: ObjectHeld , ObjectSize= d, y Obj ec tPickedUpAt Arm DEL : ClearHand

Note that we consider rigid objects so that after OPEN and before CLOSE the hand does not hold anything .

Using postconditions the ExpectedModel and the CurrentModel can be determined. For instance, after a MOVE instruction the E:~~:pectedModel is computed by updating the arm position in the lnitialModel, while the CurrentModel is computed by reading the actual arm position.

C. Knowledge base

We use a knowledge base containing two types of rules, sensor rules (used to interpret the sensor data), and recovery rules (used to produce the recovery) .

Sensor rules have the form

where the D's express what we want to know from sen­ sors and C is their "logical" interpretation. This organi­ zation allows a certain independence between the raw data from sensors and their interpretation.

The recovery rules have the form

where the G's express what we want to achieve, the S's express what we know is true, and the R's are recovery actions .

1. Examples

SensorRules:

• if FingerTouchSensorTriggered then ObjectHeld if not FingcrTouchScnsorTriggcrcd then ClearHand

Recovery Rules:

If the object is lost during the movement we can r ecover With

to obtain when do Obj cctHcld ClearHand Compute NextPickUp;

GrabObject(NextPickUp,ObjectSize)

knowing that

if ObjectPickedUpAt =X

then NcxtPickUp = X + d

If the arm is not in the right place we can use the rules

• to obtain AriTFFrarrre2 when AriTFFrrune1, Dist(Ann,Frarre2)< . 5
• do MOVE ARM TO frarrre2 DIRECTLY to obtain AriTFFrarrre2 when AriTFFrarrre1, Dist(Ann,Frarrre2)>.5 do

MOVE ARM TO f rrune2 If the hand is too closed to obtain HandOpening = ObjectSize when HandOpening < ObjectSize do OPEN HAND TO ObjectSize

2. Recovery procedure

The recovery procedure is activated by the identification of an error. As we said before, an error is identified every time CurrentModel at the end of the exe­ cution of any instruction is different from ExpectedMo­ dcl. Knowi.ng the situation in which we arc and where we want to be the appropriate error recovery rules can be fired .

We control both the preconditions before executing any instruction and the postconditi.ons at the end. The first check should not be needed since we assume that the program does not have logic errors. We consider it useful as a protective measure.

After the recovery we resume the execution of the original program at the point where it was suspended. The problem of deciding whether to restart it at a different point has not yet been approached.

III CONCLUDING REMARKS

Although the examples shown are limited ve think to have supported our claim that we have presented a general framework for error recovery in robot programs. Research is under way to write more rules, to introduce strategies in recovery. and to extend our work to com­ plete programming languages. A preliminary irrplemen­ tation is under development. ·

In our opinion the strong points of our method are:

• It is based on the use of a real robot programming language, not a planning system intended for purposes other than manipulator control;
• The reasoning process used in error recovery is based on information provided by sensors. Any sensa- can be incorporated, provi.ded that interpretation rules are available;
• The knowledge base can be easily extended to cover more errors and more recovery procedures;
• The language used to program the robot could be changed, provi.ded that its semantics is supplied in the same form;
• It can be used to recover errors not only for robots but also for more complex automation systems .