The paper (first?) introduces an idea of motion synthesis by example, i.e. producing a specified motion by combining a set of adjusted sample motions stored in a database. Very inspiring ideas and questions are discussed concerning motion database creation and usage. The approach is illustrated on a very simple model of "Luxo", hopping across a variable terrain.

The same dance motions captured with different styles (e.g. bright, gloomy) are interpolated or extrapolated to achieve style exaggeration. The joint angles are transformed by Fourier transform and interpolation (extrapolation) is performed in Fourier spectrum. The final motion is then produced by inverse FT on interpolated signal.

Synthesis of a free human motion from a database of various motion clips. The motion specified by constraints to satisfy is generated by cutting and pasting portions of the clips from the database. The portions are selected to have very similar last and first frame and the small discontinuities around the joint are smoothed. To find appropriate portions of the clips is a graph search problem. A pyramidal hierarchy of simplified graphs and successive mutations of searched path are used to speed up the search to interactive rates.
Only the motions available in the original dataset are possible to synthesize. However, also a similar motions synthesis would be very useful. Anyway, the key issue is the database choice, to be compact and allow the widest range of human motion.

The Fourier image is used to represent a periodic motion and to generate its variations. Motion transitions and exaggeration are obtained by interpolation and extrapolation; kinematic characteristics (e.g. speed, step length) are altered by adjustment of Fourier coefficients and the most interesting – difference of (Fourier images of) a "normal" motion and the same kind of motion with a mood gives the mood characteristics itself. Adding this mood characteristics to a different kind of motion the mood is transfered to a new motion.

The method is a more realistic substitution of inverse kinematics based on interpolation of several stored poses surrounding a goal pose. The stored poses are recorded by MC and regularly resampled. One of the earliest works about motion synthesis from examples.

Motion control by specifying constraints and a goal function followed by solving a big optimization problem over the whole time interval of the motion. Initial, final and "style" constraints as well as the goal function to be minimized (e.g. muscle energy) are specified by an animator, further constraints follow from physical laws of mechanics. The optimization is solved numerically. Example of jumping Luxo is presented.

Producing of motion by genetic algorithms. The most successful behaviours are selected from populations simulated in physical environment.

Simulation of 3 particular human motions (running, bicycling and vaulting) by dynamics. Any new motions would require specific analysis and respective equations formulation.

Human motion is produced by applying sine-wave signals to joint angles. Moreover a noise signal is added to avoid absolutely repetitious motion. Motion transitions are performed by interpolation with variable S-shaped weights.

Creation of transitions between last frame of a motion clip and a different first frame of a successive clip. The transition is made by spacetime optimization over the entire time interval of the transition. A term for inverse kinematics constraints (avoids e.g. support foot sliding during transition) and a term for metabolic energy (causes the motion to look natural) are both minimized. That presents an inverse dynamics problem, not very clear to me. Also transitions with motions defined only on a part of body are possible.

Motion synthesis by interpolation of example motions. Example motions of one kind (e.g. walk) but with a different style (e.g. happy, downhill) are recorded and parametrized by hand, i.e. a corresponding vector which components are values of happiness, slope, etc. is specified. To synthesize a new motion corresponding to a given parameter vector all DOFs of example motions are interpolated by linear and radial basis functions (scattered interpolation of irregular data). Parameter vector can vary over time as well as transitions between different kinds of motions are possible forming a motion graph. The realistic interpolated motion if generated in realtime.

Inverse kinematics by interpolation of examples. Scattered example poses are interpolated by radial basis functions. RBF interpolation is improved comparing to paper Verbs & Adverbs of the same authors. Weights for linear combination of examples as a whole are computed using RBF against separate interpolation of DOF's curve control points in Verb & Adverbs. Further techniques are proposed to compensate serious inaccuracies of the used kind of RBF interpolation.

A survey on scattered data interpolation in higher dimensions. The most useful to motion interpolation seemes to be the point schemes (Shepard's and radial basis) as they do not require a tesselation of the multidimensional space.

A non-uniform B-spline curve is fitted to each DOF of MC data by least squares first. Then by curve manipulations simple operations on the motion are performed yielding naive motion blending/interpolation.

MC data are filtered and fitted by a non-uniform B-spline for each DOF. Simple heuristics for a knot vector are proposed. Some usual operations (path editing, motion warping, interpolation and exaggerations) are performed on motion sequences in B-spline functional representation by a simple naive way.

Construction of a motion between given two keyframes using a motion database. Body poses from example motions are quantized (by clustering) and a Markov chain is build as well as Hidden Markov model with motion examples in database as hidden states. A new motion is synthesized by finding a path through Markov chain states and using HMM to find corresponding parts of example motion sequences.

A generic model of a motion (as a state machine) and stylistic degrees of freedom are found from examples of the motion in different styles. That all is accomplished in a fully automatic way by methods of learning (hidden Markov models, cross-entropy minimization, annealing). No manual segmentation, alignment or annotation is needed. The motion can be then resynthesized for any value of a style variable v. Results are very strong and applicable, e.g. altering the style of a choreography, enhancing the style of an unskilled performer, exaggeration, motion analogies or creation of a new complex and valid choreography.

Motion of a body part is added to a full body motion. The partial motion doesn't just overlay a subset of body DOFs' trajectories but affects possibly the entire motion. The partial motion is extracted as a difference of a base motion with and without a desired action. The extracted motions across various bases are collected and non-zero DOFs are identified. Each DOF with a small variance across the partial motions is averaged, the other ones are stored for each base motion. In motion synthesis the partial motion is added to a base motion.
A drawback: The implementation of the proposed method is strongly dependent on a human assistance and decisions made by observations. A more automatic and general variation of the method is needed.

Interpolation of cyclic motions. A sequence of significant events (e.g. foot strikes or arm swing extremities) and their respective normalized keytimes is (semiautomatically?) built for each source clip. Then a linear (for 2 source clips) or bilinear (for 4 source clips) interpolation is performed using the time warped by keytimes of events. The keytimes are interpolated first and a relative time in an interval between successive events is computed. The source motions are evaluated using the relative time in interpolation. Moreover the upper and lower body halves are decoupled – interpolated separately and their coordination is smoothly corrected... Not so clearly explained.

Improved previous paper. Interpolation of 2 or 4 motion sequences weighted by a user specified B-spline curve. Keytimes – times of important events in the motion sequence (e.g. end-effector interactions with environment or joint angle extremities) are (semi?)automatically identified and matched among example motions. The time is warped then to interpolate the keytimes consistently. Interpolated pairs or quadruples of motions sharing at least one cyclic motion may be chained. Time warping is used as a common technique to synchronize between upper and lower body timing or adjustments at joints between successive example motions. Slight feet sliding is corrected by IK. Application of the method for motion cyclifiaction is proposed.

Synthesis of motion controlled by stepping on a pressure sensor pad and using a motion library. Realtime, but with 1 second lag. Simple analytic IK.

Fixing of skating feet in a walking motion. Analytic IK and smooth blending of corrections to the original motion is used.

Human motion synthesis using a database of motion captured, unlabeled motions. The database is preprocessed – all possible transitions with a respective probability are assigned to each frame thus forming a statistical (Markov) model. Another higher level model is constructed by clustering for efficient search in the graph.
Only the original captured motions from the database can be applied to an avatar, the motions are not modified in any way.
Nice, structured bibliography of related papers.

Motion synthesis from a given footprints sequence and their respective timing information. The motion is generated in a physics-based manner with a comfort term. Trajectory of a centre of mass is found by global optimization. The leg poses are determined then by IK. The method provides valuable anticipation and recovery in the produced motion and runs in interactive rates.

Multidimensional interpolation of shapes using radial basis functions. Also reparametrization of the abstract parameter space by inserting of pseudoexamples is treated to prevent fatal deformation and cracks in the shape.