馃攳
Examining Profiles for Robotic Risk Assessment: Does a Robot's Approach to Risk Affect User Trust? - YouTube
Channel: ACM SIGCHI
[6]
Hi everyone I'm Tom and I'll be giving you an overview of our paper entitled
[11]
"Examining Profiles for Robotic Risk Assessment: Does a Robots Approach to
[15]
Risk Affect User Trust? So what were the aims of our research? We were looking to
[21]
see how a robot's approach to risk would affect a user's trust in that robot and the
[26]
reason we wanted to do this is that you might not always be able to supervise a
[28]
robot you might want to leave it unattended and make sure that it's doing
[32]
its tasks and trust that it will finish the task it was set which really
[36]
allows you to utilize its autonomous capabilities. On top of that it allows
[42]
you to have more faith in the robots information that is producing and giving
[47]
to you for example in an evacuation situation a robot may have more
[51]
information about the state of affairs than you do and you want to ensure that
[56]
a user will trust that robot and follow its instruction if its
[59]
instruction is correct. So we were looking at a human approach to risk, a
[64]
risk seeking approach, a risk-averse approach and the risk neutral approach
[68]
which led to two hypotheses our first one was that a robot that's using human
[73]
approach to risk is likely to be more trusted and that a robot utilizing a
[78]
risk seeking approach to risk is less trusted. So how do we define a human
[83]
approach to risk? We look at the literature and we find that prospect
[88]
theory seems to be the most common approach that is where risks are given
[94]
in the form of the prospect a value probability pair like the one shown on the
[98]
screen so in this case the one prospect so this
[103]
is a prospect one choice is 50% chance to gain a thousand pounds and 50% chance
[108]
to get no pounds or a hundred percent chance to gain four hundred fifty pounds
[112]
now what prospect theory says is that in the face of a potential gain like this
[117]
one people tend to be more risk-averse whereas if you flip the signs and it was
[121]
chance of losing a thousand pounds or losing four hundred and fifty pounds
[125]
people would be more risk seeking we can look at this by changing the value of
[132]
the prospect and changing the probability to create
[135]
an overall value for the prospect so the value function takes account for the
[141]
fact that people from have different opinions on the value so if in the
[146]
previous example we had a thousand pounds someone with millions of pounds
[149]
in their bank account is likely to see that as less if again and might just take the
[153]
risk for the fun of it whereas someone with a lot less money
[155]
might want to take the certain option to make sure that they gain the money in
[159]
terms of the weighting function this accounts for the fact that people tend
[163]
to overweight small probabilities and underweight higher ones
[167]
so here we have the analytical form and the interesting note about this is all
[174]
of the parameters in it are unique to each individual the way we determined the
[177]
parameters was we use the values derived in first Tversky and Kahneman's paper
[183]
So we tried to design user studies that would test which profile was more trusted
[191]
which profile out of risk-averse where the robot won't take any risks
[196]
risk seeking whether robot will always try to maximize its reward by taking risks
[200]
risk neutral where the value is based on the expected value and the human approach to
[206]
risk where it is based on prospect theory as we've already discussed
[210]
we used the unity games engine to create an ecologically plausible environment so
[215]
that people felt immersed and the task was set in a nuclear facility so people
[221]
were trying to maximize the area that the robot saw while minimizing the radiation
[226]
so here we have the experimental setup so you pass through ten of these types
[232]
of choices these gates where on one side of the screen will have one choice and
[236]
on the other you'll have another so in this case the robot is indicating by
[240]
highlighting the box blue and putting a picture of itself that it wants to take
[243]
the 50-50 chance the user can question the robot and
[248]
question the robot further if you question the robot it will give the
[253]
values it used to make its decision and when you question it further it will explain how
[256]
it came to those values ultimately the decision must be altered or accepted
[261]
if you accept the robot's decision it will take the choice that's highlighted blue
[264]
and if you alter it it will take other one so how are we measuring trust
[269]
well at each gate we were measuring whether or not people questioned the
[274]
robot or intervened with this decision with the assumption that if a robot was
[278]
questioning more it would be less trusted and if it's decisions were
[281]
rejected more would also be less trusted at the end of each risk profile we asked
[287]
users to take a questionnaire to determine their subjective level of
[290]
trust in that robot and then after they had seen all the profiles we asked
[295]
them to rank the robots in order of trust now interestingly the robot
[299]
ranking was the only measure taken after the users have seen their score for the
[303]
round and so we hoped it would help us determine whether or performance
[306]
affected trust an important note is that we decided to gamify the study so
[312]
that we could artificially introduce risk to participants so in each round the
[317]
score from each of the choices was totaled and then after all four rounds
[322]
those were tallied up users could then partake in a bonus round where the first
[327]
ranked robot in order of trust would act autonomously and that bonus round score
[332]
was added on too and then the users were compensated based on their score so to
[337]
make this clearer I'll just give a summary of the procedure
[340]
so first participants would have to take a demographic questionnaire and then for
[346]
each of the risk profiles they would pass through ten gates at the end of
[349]
each profile they would take the trust questionnaire and then they'd be given
[353]
their score at the end of all four profiles they would rank the robots in
[358]
order of trust and then the highest ranked robot perform autonomously
[362]
through ten more gates and then the score from those rounds and the bonus
[366]
round would be summed and that's what users would be paid based upon
[371]
so our results first we found that with questioning and intervention that there
[377]
was no significant difference between the level of questioning in each profile
[380]
this suggests to us that the level of questioning of a robot probably isn't a
[384]
good metric for measuring trust however we did find there was a significant
[388]
difference between the level of acceptance in each profile with risk
[391]
seeking being significantly less this suggests that the level of
[396]
acceptance is a good metric from the trust questionnaires we found
[401]
that there was a significant difference between the risk seeking profile and
[405]
all the other profiles this again suggests that a risk seeking robot is
[410]
significantly less trusted than the rest the other interesting thing that we
[414]
found was the knowledge of expected value affected their trust scores in the
[419]
questionnaire particularly it meant that participants with knowledge of expected
[424]
value rated the human approach lower and we think this is because from feedback
[432]
that they believe that expected value is the way a robot should approach risk and
[437]
it's interesting to note we deliberately recruited from a technical background as
[441]
our end goal was for a nuclear environment so our users tended to have a higher
[446]
knowledge of expected value than you would find in the general populace so if
[452]
you were maybe looking to use this in something like assisted living where a
[455]
knowledge of expected value is less it might be that the human approach to risk
[458]
scored higher finally the ranking so we found that the only significant
[464]
difference in the ranking was between the neutral and the risk seeking profile
[468]
as you can see here the risk-neutral profile received the most ranks
[472]
for first place and the risk seeking received the most for second interestingly
[476]
they suggested the performance didn't affect score and some participants even
[480]
stated that they found this was the case so what does this mean it means that
[487]
when you are trying to design a system and you want to maximise trust you
[492]
should try not to use a receiving profile which seems obvious but now
[496]
we've managed to empirically show it and also if you're designing for use by
[503]
technical personnel it doesn't matter which profile you choose between
[506]
risk-averse risk neutral and a human approach to risk you can choose
[511]
whichever best suits the implementation and also it suggests that if you're
[515]
designing for non-technical users there's the potential that a human
[519]
approach might be more trusted and be a good choice so just to conclude we found
[525]
that our first hypothesis was incorrect people did not trust a human approach
[530]
to risk more than the other approaches we found that our second
[534]
hypothesis was true that the robot using a risk seeking approach will be less
[538]
trusted we found that performance did not affect trust we found that
[542]
acceptance of a robot decision did in fact correlate with other trust metrics
[545]
and therefore could form a useful measure of trust in the future and
[549]
finally we found that questioning a robot's decision did not form a good
[553]
measure of trust so in the future we're hoping to investigate the parameters
[557]
that define the human approach so that we can tailor it to the robotic implementation
[560]
thank you for listening
Most Recent Videos:
You can go back to the homepage right here: Homepage