No, but they increase heat, energy use, weight, etc. Then you need more batteries, better cooling, bigger motors to carry the weight, etc. To do robust pose estimation of arbitrary objects (cups, books, shirts, crumpled up pieces of really important paper, etc) is an unsolved problem, and current solutions require very expensive sensors. In particular, to accurately grab something requires extremely good localisation, which means you need a truly robust SLAM implementation, or more likely, a better IMU. Getting 2-5X better in IMU terms means roughly two orders of magnitude in price.
I absolutely want to build a mobile manipulator robot one day, but I'm 99.9% convinced that it shouldn't be the first thing I build.
So it seems it breaks down into 2 distinct issues :
- cost of the arm mechanics
- expensive sensors because the algorithms aren't yet good enough to do what's needed with cheap sensors
I don't know much about the sensor issues but I'm unaware of any fundamental reason why an arm which basically consists of some metal supports, joints and a handful of smallish motors should cost more than a new car.
As for the algorithm issues - humans seem to be able to make do with 2 eyes which seem about equivalent to a pair of cameras which aren't expensive at all. I take that as a proof of principle that the problem is solvable. Looking at how much progress has been made on robotic cars since the first DARPA challenge my bet is that a well focused effort could bring us close to a practical CV based solution within a couple of years.
A lot of the cost of a Barrett WAM arm (for example, they're popular in industry and research) is the controllers and motors. In order to get sufficient precision for pick and place you either move very slowly, or fast with extremely high gains. High gains mean overshoot unless you can also run the control system at warp speed. All that combined means the solution is non-trivial, and the few companies that have solved it can sell it at the sorts of prices we quoted.
As for computer vision, sure, it's possible. But an RGBD camera (e.g. Kinect) will be better. Condensing the explanation, these sensors project a grid of near-IR light using a laser, then use an IR sensitive camera to pick up the grid, noting that it is distorted by obstacles in the environment. These distortions can be used to determine the distance to corners in the grid. A normal RGB camera picks up colour, and putting it all together we get a 3D coloured image from a single frame. You cannot do that with a single camera in a single frame. You can do it with two cameras, but arguably its harder and more expensive.
I agree that using a Kinect is probably a good early approach. 3D CV certainly isn't easy. If I were developing a product though I'd want to be careful about being dependent on one company's technology. I'm not sure how easy it would be to replicate a Kinect like device (whether for legal or technical reasons) and my impression is that high speed laser scanners are very expensive.
If you'd like to continue this discussion my e-mail address is in my HN profile. I too am very interested in developing a robotics product. One of the things I'm really trying to understand is what exactly makes these arms so expensive and what, if anything can be done about it. You mentioned motors and controllers. I wonder which one drives the cost. With motors the problem might be materials and manufacturing cost, if its the controllers then I think the cost would be more in non-recurrent engineering (because electronics manufacturing usually isn't very expensive) and that can be minimized if you have a large market.
Processor cores aren't that expensive.