Standard, exact techniques based on likelihood maximization are available f
or learning Auto-Regressive Process models of dynamical processes. The unce
rtainty of observations obtained from real sensors means that dynamics can
be observed only approximately. Learning can still be achieved via "EM-K"-E
xpectation-Maximization (EM) based on Kalman Filtering. This cannot handle
more complex dynamics, however, involving multiple classes of motion. A pro
blem arises also in the case of dynamical processes observed visually: back
ground clutter arising for example, in camouflage, produces non-Gaussian ob
servation noise. Even with a single dynamical class, non-Gaussian observati
ons put the learning problem beyond the scope of EM-K. For those cases, we
show here how "EM-C"-based on the CONDENSATION algorithm which propagates r
andom "particle-sets," can solve the learning problem. Here, learning in cl
utter is studied experimentally using visual observations of a hand moving
over a desktop. The resulting learned dynamical model is shown to have cons
iderable predictive value: When used as a prior for estimation of motion, t
he burden of computation in visual observation is significantly reduced. Mu
lticlass dynamics are studied via visually observed juggling; plausible dyn
amical models have been found to emerge from the learning process, and accu
rate classification of motion has resulted. In practice, EM-C learning is c
omputationally burdensome and the paper concludes with some discussion of c
omputational complexity.