In particular, improved performance, other forms of

6. 1

Future Work

A number of items remain for future work.

locomotion and more natural looking motion remain to be addressed. As well, extension of the

control formulation to non-periodic motions and the possibility of automating much of the design

process stand out as worthwhile avenues to pursue.

6. 1. 1

Better Discrete System Models

One of the drawbacks of the current control approach is the high computational expense of

reconstructing the discrete system model each cycle.

In the case of our bipedal control, this

results in a four-fold increase in the time required to generate the final motion.

Two reasonable

possibilities exist to reduce this cost. Both are based on the reuse of previously computed models

rather than blind reconstruction of the model each cycle.

One approach would be to reconstruct the model only when necessary.

Once a reasonable limit

cycle has been reached, the model parameters determined through sampling remain relatively

constant from one cycle to the next. In such cases, a fixed model may be sufficient. By assuming

fixed model parameters and monitoring the final RV values for deviations from the limit cycle, it

seems likely that direct balance control can be achieved for much of the desired motion.

When a

limit cycle terminates, for example due to a change in base PCG, a new system model could be

constructed.

A second approach might be to construct a general discrete system model which is parameterized

with respect to the creature's initial state at the start of a cycle. Such a model could be constructed

by generating a number of walks

from

various

initial

conditions

and

recording

the

model

parameters and initial state for each. Once a large enough number of models have been generated,

they could be used in the form of a lookup table.

Particularmodels could be chosen using a

nearest neighbour approach based on the initial state of the current cycle. The number of different