@Robotistry @clew @futurebird @mxchara
Yeah, we suffer hard from the software model of independent boxes that abstract complexity, while robotics is the opposite of that at every point.

(Also, for sensor artifacts, nobody ever has produced the special hell that is automotive radars. I'm wildly pro-radar but those things are designed to make engineers fight.)

My personal dream is for us to abandon humanoid/vertebrate-mimic body plans _and_ sharing floor real estate with humans and just start building ceiling squids. All humans will love tentacles coming out of the ceiling to help them!

@clew @Robotistry @futurebird @mxchara
(To be clear, I love robotics, it is the best job, even better than crazy radio shit)

The problem is that any fix requires fixing multiple levels at once, and those levels are in different, often warring disciplines.

For instance, if you find a vendor underspecced a motor brake (they always do) now you need to reduce the length of limbs, the available torques in software, the available currents in power management, oh right that changes which constraints bind the constraint solver so I hope the controller software has hopped on to renormalizing Jacobians an' shit, also wake up the contract lawyer, negotiate building access for the vendor's technicians, and probably 3D-print some little stop widgets as well to clip onto the motor to backstop any broken brake springs, but maybe those can wait until overnight. Also the new constants for the balance controller to avoid stressing that motor cause the robot to make lots of quick stomping steps, annoying the tenants below your lab because the vibration aerated their anaerobes or something, so your landlord is also on the phone now.

Realistically no individual person in this job _can_ be a specialist in only one area; everyone needs to know enough of everything to at least talk to each other. That sort of hyper-generalist workplace is an absolute trip to work in. And it isn't going to get solved by "AI" in short-to-medium time, although several parts of it will become moderately simpler or cheaper.

@Robotistry @futurebird @mxchara
Go right ahead

@clew @Robotistry @futurebird @mxchara
Say I want to climb a stair. Now I need a leg (mech e), which needs a balancing controller (classical control), which means estimating payload inertias (learned control), and I need a big battery (materials science) or a passive dynamic walker design (mech e again) and regenerative motor braking (ee).
Say i dig down on regenerative braking. Now I need a high-rate controller (sw. eng), low-backlash geartrain (mech e), high-rate battery charging (f'ing solid state physics) and a plan for overcharge (safety enge, ee). That plan for overcharge has to include emergency stop (classical control) of a possibly dynamically unsafe system (learned control? probably?) while on uneven terrain (mech e) which is.... back to where we started trying to climb a stair.
This is an example of a _mostly solved_ problem. The real world of unsolved problems is much, much worse. Changing any constant by a millimeter blows up all your assumptions everywhere in the stack in a way that no single engineer can have in their head at once.

Plantigrade humanoids arguably are the worst case of this problem, but autonomous cars have a few notable strange fractal doom loops where nobody can understand it all at once.

@futurebird @mxchara
So much this. The only problem in robotics is everything, because each problem in robotics is every other problem in robotics in a trenchcoat.