Shadow Robot’s three-fingered hand is robust enough for reinforcement learning | Hacker News Reader

Shadow Robot’s three-fingered hand is robust enough for reinforcement learning

spectrum.ieee.org

35 pointsT-A13 comments

> Shadow Robot

From The good-looking textured light-sourced bouncy fun smart and stretchy page[1]!

[1]: https://web.archive.org/web/20160521081256/http://freespace....

> Shadow Robot

Actually that's the prosthetic for "Three Finger Mickey" who has the greasy food window just left of the dinosaur. You can see 3FM actually delivering a dripping food sac to a customer.

[1] http://assets.rootsvinylguide.com/pictures/5-firesign-theate...

Looks like it could deliver a beating, too.

That thing could deliver a punch or get a good grip on a baseball bat. Just the thing to hook up to a machine learning system.

One day they can play instruments too which will be nice

Nothing beats a physical instrument producing sound by strumming or piano hammers striking or wood hitting drums combined with the natural reverberations of the room. Much effort has been put into "physically" emulating it in software already which of course is only ever going to be 99% of the way.

That was done long ago. See "Orchestron", "American Photoplayer", and "Mills Violano Virtuoso".

That is not the same at all. Those are very mechanical and boring.

Using human instruments with refined and subtlety random velocity and reacting real-time to whats being played is what's interesting IMO.

Ah, you want a reproducing piano. That's the really good variant of a player piano, where the expression is encoded along with the notes. They're rare; few have survived, and they were high maintenance. Here's a good one, playing back a performance by Sergei Prokofiev.[1]

[1] https://www.youtube.com/watch?v=9uMRD2o6dJo

When it comes to training, software, does physical scale really matter?

If you want to train control software for a delicate humanoid hand, why not scale it up two or three times to get robustness, and then have it manipulate equivalently enlarged test objects?

I suppose there might be some square-cube law stuff when it comes to catching and throwing...

The issue is at the early stages of reinforcement learning the system will send essentially random inputs to the motors and it's important that this flailing doesn't break the machine so that learning can continue.

Or do the bulk of the training in a virtual world, only going to the physical world for the fine tuning.

Unfortunately, noone has simulations that do soft contact well enough for this to be any good at all (which is why you see plenty of academic papers about it, but very little in terms of actual products doing it.)

This reads very much of https://xkcd.com/793/:

> You're trying to predict the behavior of <complicated system>? Just model it as a <simple object>, and then add some secondary terms to account for <complications I just thought of>

> Physicist: Easy, right?

> Physicist: So, why does <your field> need a whole journal, anyway?

> Liberal-arts majors may be annoying sometimes, but there's nothing more obnoxious than a physicist first encountering a new subject.

Yes, physical scale absolutely matters. I'm a controls engineer in industrial robotics. Some of my material handling projects involve giant Fanuc R-2000 robots that move very much like gigantic, dextrous fingers as they use incredible levels of power with their rigid gearboxes and heavy castings to manipulate payloads weighing hundred-kilogram payloads. They're fast for what they are, but they move with all the urgency of a barge compared to human fingers. They're completely different from tiny high-speed conveyor systems using little SMAC linear motors and laser galvos that are flapping about at speeds the eye cannot hope to track.

There is square-cube law stuff everywhere. Also, power density, strength-to-weight ratios, and stiffness of actuators and limbs has enormous variance - your intuition about what a mechanical gripper might do or how rigid a finger vs. a block of steel might be will likely mislead you.

Even though it might help to start with a mechanism at a different physical scale, changing the scale matters indeed.

The mass being different, this leads to different forces required to get the same acceleration or deceleration. That follows directly from the formula: force = mass * acceleration.

Moreover, the resonance frequency [1] of an object is different for different scales. For example, piano strings will vibrate at different speeds, depending on their sizes. That would be the same for the robot cables.

Also, when changing the scale of an object, the materials often change. The cables and the motors would probably be built differently, with different physical properties.

In general, to design a new mechanism, it is common practice to start with software simulations, 3D prints, and small scale prototypes. However, the properties of the new mechanism are only validated when you have built a real size prototype.

[1]: https://en.wikipedia.org/wiki/Mechanical_resonance

Reply on news.ycombinator.com