All the technology is falling into place. The navigation and vision from the DARPA Grand Challenge,
No offense to DARPA, not all of the navigation and vision algorithms in those cars with a whole set of high speed computers are really practical for use on smaller home service robots. For one, they use mostly laser scanners fused with GPS data, not practical for home. The problems in home environments which are less dynamic than Darpa'esque environments, is that they typically require precision sensor technology and a whole lot more development in data fusion and in uncertain environment decision making. Conventional sensors for the home (ultrasonic, IR, vision etc.) all have limitations heavily fscking up the performance of the control algorithms. Developments are being made in this area though.
the success of the newer algorithms in machine learning,
What success? I do admit lots of cool stuff has been done recently. But machine learning is facing some strong limitations when compared with the abilities of biological systems in coping with unsupervised learning in uncertain and dynamic environments. Structured environments like searching the web are totally different to navigating a randomly cluttered room with your sensors at 50% efficiency (try navigating a typical bar while being drunk). Newer results in adaptive connectionist systems with evolutionary optimisation has some promise though. But neural networks have come a long way with not much resulting yet.
the balance and slip control of Big Dog,
...applies to quadrupeds with the similar mechanical characteristics. If you are trying to imply that the results are relevant to humanoids, I suggest you read up on the loads of material on everything from 3d linear inverted pendulum model to spin angular mommentum regulation and control for humanoids. The fundamental difference between quadrupeds and humanoids are those of humanoids frequently not having the freedom of undertaking static walking. Although I am curious to find out how they managed to stabilise the Big Dog from the kick scenario where it definitely loses static stability (Center of mass leaving the foot polygon).
and the cost structure of the toy industry
Irrelevant to the service robot industry. The development of advanced homeservice robots have few things in common with the toy industry. Except for trivialities like servos and interesting behavior control techniques like subsumption used in most toys, home robots share few parts in the same scale or scope as toys.
We have not yet seen the "killer app", but I think that robotics is now where personal computers were in about 1976, after the Apple I but before the Apple II.
This is spot on though. I personally feel the killer app will be precipitated by a combination of realisation of cost of human labour for mundane tasks along with the economics of the usage. I was until recently in a Korean Robotics R&D, where we typically calculated a vacuum cleaning robot to prove cost effective in 6 months of usage (initial cost of 400$ + power Compared with frequent manual vacuuming) for a typical apartment. This sort of calculation when tied in with mass manufacture shall break the killer app barrier.