Hacker News Reader

Comment by gisely | Hacker News Reader

I would really love to hear why performance specs you mention are still necessary for robotics applications. Repeatability and speed certainly seem useful industrial robotics applications, but low cost also seems critical to enabling broader use of robots. I am particularly interested in the possibility that modern algorithms (e.g. deep reinforcement learning) can compensate for not having completely repeatable behaviors from a robot by using sensory feedback to create behaviors that reliable despite their variability. Additionally, materials with higher elasticity (e.g. plastics instead of metal) might allow some previous impossible efficiencies, e.g. by storing potential energy in the links and then releasing it at the right time. Is it just too soon for these sorts of approaches?

View on HN

Compromise on stiffness and rigidity in a servo system means the thing jitters about like a squirrel on speed. Robots need to be smooth, and there's precious little you can do in your controls to compensate for system instability. Precision comes along almost for free once you have that rigidity.

They didn't get the basics right. Special sauce on top needs to have a solid foundation.

And that's why I have an Aubo robot in my conference room today, and why I was able to tune in a 250kg payload 1995-vintage ABB on Tuesday, while these guys are shutting down.

> and why I was able to tune in a 250kg payload 1995-vintage ABB on Tuesday

What does that mean?

ABB is a major supplier of industrial robots, so I assume the poster meant installing/setting up a used 1995 model capable of handling 250kg.

Thanks!

Or tune in a 250kg payload on such a robot - I can imagine the motion programming might have to be tweaked to swing right. I don't know what such a heave payload might be - maybe some heavy part to be welded.

By 'tune in' he means calibrate an end effector to be mounted on the robot. This involves TCP calibration (tool center point) and specifying the centre of mass and moments of inertia. KUKA robots, for example, allow you to determine the TCP by manually posing the robot so that the tip of the end effector (tool) is at a fixed point. This is then repeated three more times from different poses (but the same tool point) so that the robot can solve for the 3 values that determine the tool's XYZ offset from the flange (mounting point at the end of the robot). The robot controller then uses this data when calculating inverse kinematics and supplying power to the axes.

As an existence proof, human limbs aren't stiff but are capable of building all the things, with jigs.

So better control + jigs should be able to do what we want, but maybe this isn't productizable yet?

A robot is basically a programmable jig. If a robot is floppy like a human, and needs a traditional jig for precision work, then the jig is inflexible and you might as well just build a single-purpose machine to do that thing.

traditional robots need to be smooth. I doubt commoditized robots will be based on traditional servos with traditional pid control, there is plenty of room for algorithm and control improvement, not too mention the huge strides being made in soft actuators. Like a post below you said, humans are able to do it, so its a question of how to cheaply artificially do that. Just because something doesn't exist yet, doesn't mean its impossible, that's the whole point of research. As an example, the group I was a part of in school made specialized control algorithms for 3D printers that can compensate for the flexibility and dynamics of cheap/thin actuators, which nearly doubled printing speed while retaining accuracy

>> I am particularly interested in the possibility that modern algorithms (e.g. deep reinforcement learning) can compensate for not having completely repeatable behaviors from a robot by using sensory feedback to create behaviors that reliable despite their variability.

Deep Reinforcement Learning will definitely make a huge difference in robotics. Already, ~50 years after the first implementations of RL algorithms [1] robotic systems trained with (deep) RL have now become so proficient that a robotic hand has learned to manipulate a cube [2]. Always the same cube (in terms of size, weight, etc) but a cube, nonetheless! Full-blown general object manipulation cannot be more than a hundred years away, or two!

__________________

[1] I'm counting from Donald Michie's MENACE, a tick-tack-toe playing machine originally implemented in matchboxes, because of a dearth of computing hardware of sufficient power- see Rodney Brooks' excellent write-up:

https://rodneybrooks.com/forai-machine-learning-explained/

[2] OpenAI trained a robot hand to manipulate a cube with RL:

https://blog.openai.com/learning-dexterity/

Note that it's always the same cube. If they could train it to manipulate anything else, you can bet they'd be showing it off (and probably publishing in Nature or something).

Basically, yes. Deep RL has the potential to make high accuracy/repeatability unnecessary, but it's still in the research lab setting.

Reply on news.ycombinator.com