Page 20 -

Ch01-I044963.fm Page 3 Monday, August 7, 2006 11:24 AM
Page 3
11:24 AM
Ch0 1-I044963 .fm
Monday, August 7, 2006
3
The system is made to judge to which model the actual hand motion trajectory belongs from probabilistic
evaluation based on two propositions: the proposition D of stating "dangerous operation" and S of "safe operation".
"Dangerous operation" leads to a stopping operation (a brake is applied to Ihc Skill-Assist), and "sale operation" to
continuing the operation. However, if the output goes out of the pretaught pattern area, the corresponding element of
the observation symbol probability distribution becomes 0. Therefore, Dempster-Shafer (abbreviated as DS) theory
is applied. We use two distinct operation control policies, Safety-Preservation (SP) policy and Fault-Warning (FW)
policy. Operation control is carried out according to a policy corresponding to the observation result of where the
hand position lies at a time: Control judgment is made depending on which policy the third proposition X which
accepts either S or D as a frame of discernment and where the observation symbol distribution becomes zero under
both HP collision and avoidance trajectory models. In the study, this way of observation-space-dependent policy
determination is referred to as HMM-OPD.
We performed 10 operation iterations to teach data in alignment with each of the two fixed trajectories of
reaching and avoiding HP. We conducted experiments based on the teaching data to verify effectiveness of an
accident prevention method by using the operation control with the proposed HMM-OPD. In the first accident
prevention experiment, we could successfully prevent all 20 trials of the subject's motion hand movement reaching
HP from colliding with it by use of the proposed HMM-OPD method. In the other 20 collision avoidance experiment
iterations, it is judged to stop operation totally for 18 times out of 20 repetitions. Such unnecessary halts in operation
indicated that productivity might decrease severely in the event of application of HMM-OPD and also imply that
either: 1) definition of the FW and SP space was not initially optimized or 2) teaching data concerning safe operation
were in short supply which resulted in 4 times of stop operation in FW space. The following section proposes
a workability improvement process comprising renewal of both state policies and teaching data as a solution to
coping with problems remaining in the accident prevention method proposed so far.
IV. EXTENT1ON OF H M M - O P D TO WORKABILITY IMPROVEMENT
In the study,the combination of safety and productivity is referred to as workability; optimal workability is
defined as the ability to perform a task in the state where safety is secured and there is no unnecessary stoppage
of operation by the Skill-Assist. A workability improvement process is carried out in the process of repeating what
we call "hazard simulation" and HMM reconstruction is eventually performed.
Generally, it is difficult for an operator to teach initial conditions to optimize an observation space in which
a higher priority is strictly determined for either safety or productivity. Therefore, heuristic reconstruction of the
observation space is attractive. Next, we performed a hazard simulation process that is characteristic of presenting
no danger of real pinch and contributes to minimizing the volume of SP space through demonstrations in off-line
operation.
Then, HMM is reconstructed using the constructed space. Fig. 4 shows results of a collision avoidance experiment
after such a workability improvement process in which the operation is carried out along the trajectory pattern toward
a HP. In this case, because the volume of FW space is expanded by 12 times of FW-SP space-policy renewal,
the FW policy is implemented until the operator's hand reaches r = 0.09 m, when another judgment is made to
continue the operation. Moreover, it turns out that probability of dangerous operation is equal to 1 at r = 0.09 m.
Subsequently, a different judgment is made to cease operation; then accident prevention can be performed and
eventually v = 0 m/s at r = 0.06 m. This leads us to infer that the process is a useful method for optimizing the
observation space.
r=0.09[m]
r=0.09[m]
-11 1
100[%] r=0.09[m] 100-1.89x10 [%]
100-1.89x10 -1[%]
100[%] r=0.09[m]
v[m/s]
]
% 100
[ 0.4
y 80 λ 1
t φ[rad]
i hand motion trajectory
l 60 ajectory
i
b 40 0.2
a
b 20
o λ 2 0.1 r[m]
a r p o 0
0.15
0.25
0.2
0.1
0.05
0 0 0.05 0.1 0.15 0.2 0.25 0.2
r[m] r=0.06[m]
r[m]
(a) Relationship between
(a)Relationship between
distance and probability
distance and probability
(b)Hand motion trajectory
(b)Hand motion trajectory
Fig. 4. Stop operation for collision avoidance after FW-SP space-policy renewal
The second renewal process of teaching data is also implemented successfully, and collision avoidance experi-
ments demonstrated that we could obtain a clear result of 20 continuous successful operation repetitions with no
halted operation in cither FW or SP space after workability improvement processes.