To make high computing power portable, as could be needed for machine perception applications in robotics, high power efficiency is needed. This is something the Parallella is aiming for.
Machine perception algorithms using "brain-like" computation structures are not only highly parallel by nature but also need fast access to memory. Instead of having a thin pipe to a central memory, it needs to be distributed, build close to/in to the computation cores.
The long-term vision of Parallella’s Epiphany multi-core processors indicate a 1000 cores per processor, with 128KB memory per core by 2014 a possibility(!) This is exactly the order of parallelism and distributed memory Machine perception applications need.
They understand that the key is getting developers writing software in a new way, and the trick there is ubiquitous hardware. They also understand that it has to start as a co-processor, too, so they've married it to an existing ARM chip to provide peripherals and the host operating system. These decisions are very, very good ones, especially as compared to other multicore processor manufacturers (XMOS and GreenArrays are the two I'm thinking of).