10

Badler[BPW93] uses primarily

technique has also been applied to human running [BC96].

kinematic techniques as well as rotoscoped data with dynamic enhancements to achieve many

human motions and behaviours.

Boulic, Magnenat-Thalmann and Thalmann [BMTT90] and Ko

and Badler [KB93] present techniques to generalize rotoscoped or motion captured walking data

to other subjects and step lengths while reducing or eliminating the resulting ground constraint

violations.

Raibert and Hodgins [RH91] use full dynamical simulation with robust hand-crafted hopping

control to attain various bounding gaits for biped and quadruped robot models.

As well, a

similarly controlled planar kangaroo model is shown to compare well to its real-world counterpart.

Stewart and Cremer [SC92] use their flexible constraint-based approach to generate fully dynamic

3D bipedal walking on level terrain and up a flight of stairs.

One of the required constraints,

however, is a 0 DOF "magnetic boot" on the stance foot.

Van de Panne, Fiume and Vranesic

[VFV92] use optimal state-space control tables to control walking on level terrain and up and

down ramps, smooth curved surfaces and stairs for a planar biped model. This approach requires

a suitable control decomposition to make the generation of the state-space controllers tractable.

Auslander et al. demonstrate automatic synthesis of interesting 2D bipedal walking and tumbling

motions but meet with difficulty in their initial attempts to extend this approach directly to 3D

[Aus+95].

Van de Panne and Lamouret propose the use of guiding external forces to initially

attain reasonable controllers using similar automatic synthesis [vL95]. The forces are then reduced

in a number of

steps

and

can

sometimes

be

entirely

eliminated

to

yield

a

fully-balanced,

automatically synthesized motion.

Examples of human walking, skipping and

running

and

walking over varying terrain for a simple 3D biped are given. One difficulty with this approach is

that the removal of guiding forces must be performed incrementally over the

entire

motion

sequence (for example, each step of a walk).

This process

that

can

become

prohibitively

expensive for more complex creatures.

Hodgins et al. [H+95] show how Raibert's hopping