[throwawayffffas]

I think it's about kinematics, the more precise your gears the better the model fits the real world.

That's why pro crews don't use gears and ropes. At high impulses deformations and elasticity throw the kinematics off what's actually happening. Modeling the deformations and the elasticity is a computational no no. Instead what you see is the motors right on the joints.

At least that was the case last time I had a look at robotics.

[michaelt]

> That's why pro crews don't use gears and ropes. [...] Instead what you see is the motors right on the joints.

The answer here, as with so many things in robotics, is: It Depends.

UR10e robot arm that can lift a 4kg object with a reach of 1m and has sub-1mm repeatability? Strain wave gears in the base and shoulder joints, 100:1 ratio.

MIT Mini Cheetah robot dog that can do backflips? 6:1 planetary gearbox.

Shadow Hand with 20 degrees of freedom? Tendon driven, with the 20 motors in the forearm to keep the fingers slim.

Little dinky Huggingface SO-101? Servo motors, integrating 1:345 gearing with a series of 6 tiny brass gears.

Mid-price CNC milling machine, if you call that a robot? Really long ballscrews, driven by stepper motors.

[kragen]

Surely you mean driven by closed-loop servos with encoders, don't you? Jacques Mattheij wrote a long post about how he ended up having to replace all the CNC machines he sold that used stepper motors because he didn't know any better at the time, and brushless motors are a lot faster and more powerful as well as not losing steps.

[michaelt]

I guess "mid price" isn't the most helpful term when machines vary from $100 to $100,000 huh?

I think the X-Carve 3-axis wood carver uses stepper motors with belts of all things. The Shapeoko Pro is leadscrews and stepper motors. Wazers, I believe, are belt-and-servo driven. And a lot of 'CNC conversion kits' you can order online use stepper motors. Plus of course laser cutters have really low torque requirements, so they've got a lot of design freedom.

Arguably those are "cheap" rather than "mid price" it just felt weird to declare a $4000 machine to be "cheap"

[WillAdams]

Actually, it was the Shapeoko 1 which introduces the idea of belts (MXL), then switching to Gates GT2 belts with the SO2 (ob. discl., I got a free machine for doing the instructions, see Github). Since it was opensource, the X-Carve was forked from the SO2, though the X-Carve Pro was a completely new design.

Note that belts were continued through the Shapeoko 3 (I got a machine as a "thank you"), though the Z-axis got a leadscrew in the Z-Plus upgrade, then the Pro (since, I got a job with the company and got an XXL as part of my employment), then the 4 (the original Pro is referred to as a 4 Pro sometimes), and it is only with the 5 pro that ball screws were switched to for all axes (and I now have a 5 Pro).

For an example of what a belt-drive CNC can do see:

https://community.carbide3d.com/t/hardcore-aluminum-milling-...

[kragen]

I think there are CNC machines that cost 100 times more than US$100k.

[regularfry]

Some products should be compared on a log scale. CNC machines are a very good example.

[kragen]

The middle of a log scale from US$100 to US$10M would be US$30k.

[Joel_Mckay]

In general, for some platforms each gear mechanism adds backlash precision loss, lower energy efficiency, and might not be back driven.

>Mid-price CNC milling machine

A ball-screw is mostly decorative on small machines... =3

[michaelt]

On the other hand, a 100:1 gearbox gives you much-needed torque if you're lifting a load at the end of a long arm, it makes your encoder 100x more precise (in terms of repeatability) and it makes your motor brake 100x stronger.

Back-drivability is the enemy of precision, so many robotic applications can do without it.

[Joel_Mckay]

Almost all modern servo driven units I've seen prefer to allow some compliance in the end effector. The UR5 and UR10 series for example can use force limiting control loops, and are safer to use around people.

The old "fast, cheap, or good... choose any two joke is mostly still true. =3

[michaelt]

The UR10 uses 100:1 strain wave gearing in the base and shoulder joints.

You’re right that it has a freedrive mode, and force control modes. But it’s a rigid, low-backlash robot with the compliance achieved in software afterwards.

Expensive, naturally, but none of the problems that come with things like series elastic actuators.

[topspin]

> At high impulses deformations and elasticity throw the kinematics

Sure. Yet evolution has achieved astonishing kinematics with all manner of deformation and elasticity inherent to the materials, and also constantly changing physical properties, using low resolution data. We cannot build permanently lash free mechanical devices at reasonable cost and reasonable size/weight. Eventually, the answer must be pervasive real-time compensation throughout the kinematic model.

> Modeling the deformations and the elasticity is a computational no no.

Why? Nervous systems do this. That's why you can change your shoes and still walk upright.

[BHSPitMonkey]

In the video he mentions how the specific type of Dyneema cord he's using is well-suited for the application (compared to other kinds of rope/cord). It's particularly strong, light, and inelastic; a lot of climbing equipment uses a version of it for similar reasons.

[PaulDavisThe1st]

more than 30 years ago I was writing code for a "robotic" device that used motors, directly on the joints.

the motors were so sloppy the company wasted a ton of money [0] having me write heuristics to tackle the errors they accumulated over several hours.

one of his whole points is that by using dyneema (rope), there's almost no elasticity at all in the capstans.

[0] relative to the cost of better motors

[mlhpdx]

> Modeling the deformations and the elasticity is a computational no no. Instead what you see is the motors right on the joints.

That sounds like “It’s not wrong, we just don’t do it”. There are some amazing examples of imprecise drive systems compensated for by excellent control systems all over the world, for millions of years.