A universal cellular automaton is a cellular automaton which, like a Turing machine, exhibits universality.
von Neumann proved that an automaton consisting of cells with four orthogonal neighbors
and 29 possible states would be capable of simulating a Turing
machine for some configuration of about cells (Gardner 1983, p. 227).
The outlines of a proof that the two-dimensional game of life outer-totalistic cellular automaton is universal were given by Berlekamp, Conway, and Guy (1982) and independently by Gosper (Gardner 1983, pp. 250-253). Around 2000, a Turing machine was explicitly implemented in life by P. Rendell (Rendell, Adamatzky 2001). While Rendell's machine can be made into a "true" universal computer simply by making his tape infinite, he neither noted this fact nor provided an actual construction of a universal Turing machine. Subsequently, on November 11, 2002, P. Chapman constructed a game of life pattern that implements the actions of a universal register machine, thus explicitly proving the game of life to be universal.
More amazingly still, even one-dimensional cellular automata can be universal. Wolfram (2002, pp. 644-656) gave an example of a 19-color universal one-dimensional next-nearest neighbor cellular automaton in which a block of 20 cells is used to represent each single cell in the cellular automaton being emulated. The examples above show the first few steps of the 19-color universal automaton emulating rule 90 and rule 30, respectively (Wolfram 2002, pp. 646-647).
Smith (1971) showed that 18 colors and nearest-neighbor 1-dimensional rules could be universal, and Lindgren and Nordahl (1990) constructed a 7-color nearest-neighbor universal cellular automaton. And most amazingly of all, as shown by Wolfram (2002, pp. 675-691), two colors and nearest neighbor rules are sufficient for producing universality in a 1-dimensional cellular automaton. In particular, although it is anything but straightforward to prove, the rule 110 elementary cellular automaton is universal (Cook 2004).
Gács (2001) proved that there exist fault-tolerant universal cellular automata that retain reliable simulation under sufficiently small local error probabilities satisfying the specified probabilistic noise model.