Home Creators Posts Import Register Favorites Logout
haven't archived this post yet. have a subscription? use the importer!

Content

Here's next week's video, early for Patrons. This one goes public on Tuesday 7th April. So it just about works, at least well enough I'm confident enough in the principles involved, but it looks like it's going to need a mechanical redesign as I detailed in my previous Patreon post.

Files

Previews only

Mini Robot Dog #5 - Dynamic Walking

This is part 5 of my new Mini 3D Printed Robot Dog project. This time I'm setting up variable rate interpolation so we can make a walking sequence play back at variable velocity. We can then use the IMU data to attempt to make the robot balance on diagonal legs so it walks dynamically. I've decided to put my Robot Dog, openDog, on hold for now and build a smaller prototype robot dog for testing to see if I can actually make it walk. At some point I'll return to openDog, but I believe my smaller, compliant design will provide an easier route to stability. Affiliate links - I will get some money of you use them to sign up or buy something: *************************** Matterhackers 3D printing supplies: http://www.matterhackers.com?aff=7500 Music for your YouTube videos: http://share.epidemicsound.com/xrobots Amazon US: https://www.amazon.com/?tag=xrobotsuk-20 Amazon UK: https://www.amazon.co.uk/?tag=xrobotsuk-21 Amazon FR France: https://www.amazon.fr/?tag=xrobotsuk0c-21 Amazon DE Germany: https://www.amazon.de/?tag=xrobotsuk02-21 Amazon ES Spain: https://www.amazon.es/?tag=xrobotsuk07-21 Amazon IT Italy: https://www.amazon.it/?tag=xrobotsuk08-21 *************************** You can also support me on Patreon or buy my Merchandise: *************************** Patreon: https://www.patreon.com/xrobots Merchandise: https://teespring.com/stores/james-bruton *************************** Other socials: *************************** Instagram: https://www.instagram.com/xrobotsuk Facebook: https://www.facebook.com/xrobotsuk/ Twitter: https://twitter.com/xrobotsuk *************************** CAD and Code for my projects: https://github.com/XRobots Huge thanks to my Patrons, without whom my standard of living would drastically decline. Like, inside out-Farm Foods bag decline. Plus a very special shoutout to Lulzbot, Inc who keep me in LulzBot 3D printers and support me via Patreon. HARDWARE/SOFTWARE Below you can also find a lot of the typical tools, equipment and supplies used in my projects: Lulzbot 3D Printers: http://bit.ly/2Sj6nil Lincoln Electric Welder: https://bit.ly/2Rqhqos CNC Router: https://bit.ly/2QdsNjt Ryobi Tools: http://bit.ly/2RhArcD Axminster Micro Lathe: http://bit.ly/2Sj6eeN 3D Printer Filament: http://bit.ly/2PdcdUu Soldering Iron: http://bit.ly/2DrNWDR Vectric CNC Software: http://bit.ly/2zxpZqv Why not join my community, who are mostly made up of actual geniuses. There’s a Facebook group and everything: https://www.facebook.com/groups/287089964833488/ XROBOTS Former toy designer, current YouTube maker and general robotics, electrical and mechanical engineer, I’m a fan of doing it yourself and innovation by trial and error. My channel is where I share some of my useful and not-so-useful inventions, designs and maker advice. Iron Man is my go-to cosplay, and 3D printing can solve most issues - broken bolts, missing parts, world hunger, you name it. XRobots is the community around my content where you can get in touch, share tips and advice, and more build FAQs, schematics and designs are also available.

Comments

Anonymous

Wondering what the wire ( umbilical cord ) was providing?

XRobots

That's just the programming cable - it's constrained at the robot end so it doesn't break the connector

Unogazzy

So that's how a zombie robot moves... Give me power :D Is able to walk if you straighten the legs a bit?

XRobots

There are some challenges with the geared down servos and the end stops - even if one servo is the leg can go straighter it's likely that the other one has reached the end so the leg cannot move any further in its inverse-kinematic axis. Ultimately there's a limit to what I can do with r/c servos in a robot this size, so I'm just going to do the best I can before moving to proper motors again.

Pete Mertz

Brilliant to watch, as usual! Watching the progression of walking via the IMU, I wonder if you've considered taking your model and putting it into a physics engine like Unity. From there it seems like it might be easier to iterate on quickly. The other reason I was thinking of that is that I began wondering "I wonder if you could use a neural network for this" where the inputs are the IMU data and you try to minimize wobble while still attempting to reach B from A. After all, it won't be much of a robot uprising if the robots can't think for themselves :) (Forgive my rambling... I've been spending too much time inside recently...)

XRobots

I've played around with Unity quite a bit, but it's more of a game engine, so the physics can be a bit weird sometime and it'll be hard to set all the parameters for every material so they match the real world. There are other industrial solutions, so it's hard to get the slop in joints in there and other real word factors. Learning on the real dog would be interesting, but I guess that can come once the mechanics are good, and probably with collaboration from somewhere.

Anonymous

Have you tried other servo vendors/models? Chinese servos are known to have low-quality copycat issues so buying on aliexpress you might get 1 out of 3 not work or fail early. Also, low cost motors tend not to have robust engineering data sheets, so it's always a bit of trial and error. I saw in the BOM you're using ds3120s - have you looked at tower pro mg996r/mg946r? They may have less mechanical slop and a better dead band.

Pete Mertz

Getting the necessary data out of trials would be very time consuming in real life, so it might be less of an option to test on the "real" dog. That being said, maybe there lies a hybrid solution somewhere -- i.e. using a simulation to get most of the way, and then fine tuning via transfer learning with the real physical dog. Happy to talk more about this in the future!

XRobots

Well these were advertised as that model, but they don't have any markings. I bought some with the DS3120 written on them and they are worse. I've used these servos in quite a few projects and they seemed ok for the price (£20). I've had a few MG996 servos as well and they seem about the same - neither of either has ever actually failed if I used them sensibly. I think the next logical step would be coreless/brushless servos, but then I might as well use actual brushless motors and the ODrives out of openDog - and go back to back-drivable gearboxes like my other test dogs. Quite a few decisions to make about the next steps after this design.

Anonymous

so you have a very fun crowd-interaction rig there, just split the leg functions out into two controllers per person, so one person per leg, put some outriggers with little wheels on each side so in never fully falls over then let them try to walk it...fun fun fun! ;]

Anonymous

Most robots that learn to walk (Erwin Couman's group comes to mind - Bullet Physics guy) learn entirely in simulation. However, I would recommend against this actually. I have tried this before, physics simulators are just too inaccurate (even "good" ones like ODE, Bullet, Mujoco), plus modeling it is a huge pain if you don't know the physics engine inside-out. I would recommend training directly on the robot. I have done this before, it works reasonably well. The way I did it is using a camera that looks at the ground (attached to the bottom of the robot) to detect velocity through optical flow, and then use that as a reward signal to an incremental reinforcement learning algorithm. One can code in safeguards to e.g. prevent tipping the robot. One can also use dedicated optical flow sensors, like the ones in computer mice.

ingmar meins

It looks like it's on the way home from the pub :-)

Eric lamb

Building a legged robot is not really that difficult but getting it to walk dynamically now that's tricky. I'm trying to finger that out on my biped robot and it's going to be a challenge. I look forward to seeing your progress.

Pete Mertz

That's fascinating! How many runs typically do you find it takes for the reinforcement learning to get results in real life? I've never used neural networks outside of a large dataset running on GPUs before, so I'm curious how long it take on something that's bottlenecked by a physical object!

Anonymous

It doesn't take that long, actually. Like 5 minutes to get it to "hobble", then more time to get an actual walk. More time is always better, of course. If needed, one can pause it and swap batteries. We don't train episodically, actually. Just one continuous go. Our company (Ogma Corp) specializes in very low-power learning environments, so we can run stuff on things like a Pi Zero. Our algorithms are very different from standard deep learning, though. It's for more biologically inspired, and uses sparsity as well as online/incremental learning to get away with very little compute.

Pete Mertz

Very cool! That's so much faster than I expected too. Going to look into Ogma tonight :)

Anonymous

Have you considered backlash compensation in software as is sometimes used in a CNC machine?

XRobots

Yes I considered just having some twist in the other direction using the yaw I coded in the kinematic model, but I'd rather fix the mechanics first.

Andrew Hoadley

Hi James, I'm a programmer. Wondering if you'd be interested in me helping to modularise your code to make it more reusable. for example creating a CompliantJoint class that encapsulates everything to do with joint to hide some of the complexity and make it more readable. And creating a Leg class for example, which takes three compliantjoints for input

XRobots

Thanks Andrew, I could probably do it if I really tried. It's more of a 'sketch pad' for now while I work out what I actually need to do. We'll see how things are after the new version.

Adam Purdie

I'm using the dynamixel x-series servos for my latest spiders, they're considerably more expensive than hobby servos but the difference in flexibility and power is night and day....

XRobots

I'd like to give them a go, but I'm trying to keep it accessible for now. At some point in upgrading it's almost cheaper to go back to brushless motors and ODrives again.

Adam Purdie

They do get prohibitively expensive, especially when you've got 8 legs :) But they're amazing bits of kit, position/load feedback on a 3mbit serial bus and their power to weight is bes in class.I'd like to give the odrives/brushless a go for my next large robot (humanoid sized) but i don't have your skill :) hence why my eyes are glued to all your videos all the time and usually many times on repeat :)