Papers on Low-Cost Full-Scale Mobile Manipulation

[Tony Pratkanis]

Hello All,

I have been reading up on papers on low-cost full scale mobile manipulation. There's a lot of smaller experiments such as the LeRobot and the LeKiwi, but these don't have a lot of range or height they can grasp at. I found several interesting papers.

XLeRobot - this is a mobile base by placing a Lekiwi base under an IKEA cart, then mounting LeRobot arms on top of the cart. See https://github.com/Vector-Wangel/XLeRobot. I think this is an interesting idea, but it has low payload (400 g per arm).

AhaRobot - this is move complex to build but it features elevators and a SCARA-like arm design. See: https://arxiv.org/abs/2503.10070. This layout of SCARA + elevators means that joints are not under high torque at the base of the arms. It also has a higher payload (1.5 kg per arm).

There's also Nori Bot which appears to be a hybrid of these two approaches. https://arxiv.org/html/2605.16537v1. Also payload of 400 g per arm.

For payload, I have looked into average objects that people carry on a daily basis. One of the heaviest objects appears to be a milk jug at around 4-6 kg, so I think 6 kg total payload would be a good goal to achieve.

Anyway, has anyone else seen or built anything similar to one of these?

Thanks,

Tony

[Chris Albertson]

I worked on this problem before. Buying a system will be expensive, tens of thousands of dollars, since you’re outside the hobby range. 

For this project, buy a 3D printer that prints in high-end plastic and learn to use a free 3D CAD system like Fusion360.

It’s not rocket science. Start with a flange that hands can bolt to and decide on the payload and reaching distance. Design a wrist system that can do that. The elbow must lift the wrist system, payload, and elbow. Calculate the required torques at each joint them do a detailed design from shoulder outward.

Use high-quality bearings and avoid radial loads on the motor shaft.  I can’t say it enough: “bearings matter”. if you have poor ones you get machanical slop and flex.

Below is a low-cost human scale shoulder design using off-the-shelf motors and six identical 3D printed parts. But a full gallon of milk might be too much.

Arms are easy because they don’t need the speed legs need. Legs need speed for balance or “control bandwidth” for micro-level movements. Arms don’t need that, so you can use cheap geared steppers like in my design. You can iterate and improve it. It’s human-scale and low-cost, but only the shoulders are done.

Of course, the hardware is the easy part.   Motion planning is far harder.

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/kW_fPESu56aw-56kJ1C4TTVbaqHMs0gem_0imbHt2fmnmButP5J9OhckoDFKUh5U2a75fgLZ6xQ3unGNbR_lpuiAwCBnyw8tephpIKwsEbE%3D%40pratkanis.co.

[James H Phelan]

Chris,

Nice work! Would like more details about the stepper motors and bearings you used, source, cost.  Any wisdom you can impart re selection.

James H Phelan
"Nihil est sine ratione cur potius sit quam non sit"
Leibniz

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/F891C7E6-15C1-4299-AD44-F1D902CAE95E%40gmail.com.

[Dave Everett]

Great links Tony.

I use a linear lift to get my robot arm to the ground and up to a workbench height.

Dave

--

[Ken Gregson]

In the LeRobot "line" there's also the AlohaMini

https://github.com/liyiteng/AlohaMini. It uses a SO-ARM track axis for vertical linear motion/lift

https://github.com/avenhaus/SO-ARM100-Track-Axis

I'm building a couple variants of it AlohaMini "LeDomo" to keep my LeKiwi "Midori" company. Open question on how robust a PLA gear and toothed rack will be.

The SO-ARM-200 of the AlohaMini is more capable than the -10X  Arms.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAG61pbeGitLH6sBOWjb%3DVzroQDpfbGoYF%2BuVwa3qgzh_1okHdw%40mail.gmail.com.

[Wayne Gramlich]

Greg:

That is really neat.  It is wonderful that the Aloha robot stuff has started
to migrate into robot hobbyist's robots.

I worked on the Aloha robots at Google as a contractor (OLogic) shortly before
my retirement.  Alas, my NDA is still in place, so I can only point a paper:

* ALOHA2: An Enhanced Low-Cost Hardware for Bimanual Teleoperation:
  * https://aloha-2.github.io/

Have fun,

-Wayne

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAArUYGtBQ8dhzHczP4STaJcJhq4Mk4cHr0ARkAh758J99czSig%40mail.gmail.com.

[Ken Gregson]

Wayne,

Thanks, I'm familiar with some of that work.

I think we are at an inflection point in the availability of more generally useful robots and their accessibility to hobbyist and DIY developers. 

We're hoping to build a group project around the AlohaMini in our local robotics club and develop that into a STEM course for high-school students. 

Best, 

-Ken 

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/47861fa0-eae7-48dc-bc16-0c9f5b072f9d%40gmail.com.

[Chris Albertson]

On Jun 13, 2026, at 4:05 AM, 'James H Phelan' via HomeBrew Robotics Club <hbrob...@googlegroups.com> wrote:

Nice work! Would like more details about the stepper motors and bearings you used, source, cost.  Any wisdom you can impart re selection.

This was intened for teleoperation and handing of small objects.     Only the easy part is done.     You need to select a hand.  I say “select” because there are many pretty good ones.   I like the “Yale OpenHand."  But it looks very “robot-like’.  THere are also open source prosthetic hands that look human-like and work well.   The Yale hands are far easier to control and do “grasp planning” and human hands need to be teleoperate by the human with gloves.

The link I posted goes to a model browser it s kind of clunky to use compared to a full CAD system but it will allow you to see bearing and motor details.    Try this…

Click the “flow chart” thing at the bottom now you can see a tree structure of assemblies and sub assemblies and parts.  The “eyeball” makes the visible or not.    make only one of the six link blocks visible.      Zoom in.    Next make all the 3D printed part invisible and what is left is the motor and ball bearings.   These are standard parts and the part number is the object’s title.    The bearing come out is “AST’s catalog and the Motoers are from “Stepper OnLine” and I used their part numbers.        

What is not shown in the model are detrital like threads in the holes or fasteners.  The bolt circles are generic holes not sized for clearance or threads tapping.   I think the blocks could be printed in ASA, ABS or even carbon-filled version of that.  And the other parts could be aluminum or in a prototype plastic.        The same Fusion360 design files can drive either a printer of a CNC machine.

About plastic.  I find I can use it except where there is consetraited stress.  For example a pulley that has to fit a motor D-shaft with a gib screw. The D and screw will scrip LONG before the timing belt slips a cog. SO I make metal pully hubs and press fit them into plastic timming pulleys.  They last “forever”.  Just press them in with a hammer and CA glue.    Details like this are not shown.  The moters would likey strip a plastic keyway.   That explains the stupid design where the bearing areas of the bolts is 5X larger then the key slot,   One is plastic the other is metal and metal is 5X stronger, at least.

The 3D models of the bearings and motor are from the manufacturer.   Almost everyone who sells machanical parts has links to CAD files.  Even McMaster Carr has models of every screw and nut they sell and you really have to use those models when you do a design so you are 100% sure the parts will fit.

The shoulder uses 6 identical sub assemblies for one good reason, I can build one and test it, make changes and then test it for weaks with a dummy weight attached and then after it work build five more and be pretty sure it will work.   

I am very slowly now working on a “dog-robot” that has 4 legs each made of three identical assemblies. so 12 total.  The hang-up is a moter with enough power that is light and inexpensive.  Drone moters are the way to go here.

Geared stepper moters are very heavy but for this one use case here the weaight is in the sholders and not the hands and the shoulder weight goes to wheels, weight does not matter.   I would not use stepper in a walking humanoid robot.    This robot was going to be a humanoid in a wheel chair, literally, using a moterized wheelchair as the base.  It could work in any ADA complaint wheelchair accessable space.   Then I decided I did not want such a large robot in my house and workspace. and moved to a small dog bot.   That worked well enough but because I went for cheap moters the prefromance is poor, it will never jump or run.

You will 100% need a 3D CAD system,   Eyeball-engineering does not work wit a system this complex and you have to think about things like wire routing and if a hex screwdriver can actualy reach each screw.  All before you print the first part.

Arms are easy because they don’t need the speed legs need. Legs need speed for balance or “control bandwidth” for micro-level movements. Arms don’t need that, so you can use cheap geared steppers like in my design. You can iterate and improve it. It’s human-scale and low-cost, but only the shoulders are done.

Of course, the hardware is the easy part.   Motion planning is far harder.

On Jun 13, 2026, at 12:17 AM, Tony Pratkanis <a...@pratkanis.co> wrote:

Hello All,

I have been reading up on papers on low-cost full scale mobile manipulation. There's a lot of smaller experiments such as the LeRobot and the LeKiwi, but these don't have a lot of range or height they can grasp at. I found several interesting papers.

XLeRobot - this is a mobile base by placing a Lekiwi base under an IKEA cart, then mounting LeRobot arms on top of the cart. See https://github.com/Vector-Wangel/XLeRobot. I think this is an interesting idea, but it has low payload (400 g per arm).

AhaRobot - this is move complex to build but it features elevators and a SCARA-like arm design. See: https://arxiv.org/abs/2503.10070. This layout of SCARA + elevators means that joints are not under high torque at the base of the arms. It also has a higher payload (1.5 kg per arm).

There's also Nori Bot which appears to be a hybrid of these two approaches. https://arxiv.org/html/2605.16537v1. Also payload of 400 g per arm.

For payload, I have looked into average objects that people carry on a daily basis. One of the heaviest objects appears to be a milk jug at around 4-6 kg, so I think 6 kg total payload would be a good goal to achieve.

Anyway, has anyone else seen or built anything similar to one of these?

Thanks,

Tony

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/kW_fPESu56aw-56kJ1C4TTVbaqHMs0gem_0imbHt2fmnmButP5J9OhckoDFKUh5U2a75fgLZ6xQ3unGNbR_lpuiAwCBnyw8tephpIKwsEbE%3D%40pratkanis.co.

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/F891C7E6-15C1-4299-AD44-F1D902CAE95E%40gmail.com.

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/2b866b78-2a84-4e5e-bddd-0f2da9e9018f%40hal-pc.org.

[Chris Albertson]

On Jun 13, 2026, at 7:45 AM, Ken Gregson <ken.g...@gmail.com> wrote:

Wayne,

Thanks, I'm familiar with some of that work.

I think we are at an inflection point in the availability of more generally useful robots and their accessibility to hobbyist and DIY developers. 

We can build nice puppets that must be teleoperated, and we can build what are basically CNC machine tools that run scripts for prerecorded motions.   Look at Elon Musk's robot demo at Universal Studios a while back.     The guy has nearly unlimited funds and an army of robotics engineers, and he literally had to bolt the robot’s feet to the floor so that robot dancers did not fall over.   Yes, fixed-feet screwed to the floor.      His other robots could walk on flat surfaces but not step up or down a curb and still needed full remote control.     And THAT is with a billion-dollar budget.

Even Boston Dynamics’ “Atlas” that can run up a ramp and do a backflip while jumping off a platform.     That was all programmed using motion capture for a human performer.   If you were to move the ramp two feet, Atlas would run into a wall and fall down.   

I’m not saying it was not hard; Atlas had to do a ton of computation to reach the motion targets that were recorded.  It was MPC done in a server room; it was a huge optimization and search problem.   But in the end, it was a dancer who invented the performance.  The robot was a puppet.   This is the best that the best organizations can do in 2026.

We are genetically programmed to be fooled.  As social animals, we use our own brains to predict what others might do.  When we see a human, we assume they are like we are and operate the same way.  Our brains place us in their shoes and predict what we would do.    We use this without thinking about it.      We even use this to place ourselves in different situations and figure out what we would do if things were different.         This is not uniquely human; monkeys and chimpanzees and dolphins all do that some thing.

So when we build a robot that looks human, we trick people into thinking that the robot walked across the room because it wanted to be on the other side.  Or with CPT-chat, we think it “figured out” a problem like we would.   Some research showed that simply placing a happy face sticker on a robot enhanced this instinctive social animal trick.    This effect makes us think robots are dramatically more advanced than they really are.  We can’t be “un-fooled” 20 million years of evolution burn this into our DNA.

I don’t think we are very far along; we can make animated mannequins that are as smart as a toaster oven.   I’m not saying that as a criticism, but rather a challenge. The future is wide open, and one smart guy can still make a revolutionary discovery.   AI researchers in 2026 are like physics was before Isaac Newton.   The basics are still to be discovered.   It is good to work in this field.

But robots do make a good subject in a classroom.  Students need to learn how to solve problems like “How many M4 screws, in single shear,are needed to transmit 700 cm kg torque using a 75mm bolt circle?”   Or what rotation sensor accuracy is needed to give 1mm precision at the end effector or 100 other simple robotics engineering problems.   Those are both good high school level problems covering static cases.        Movement quickly gets harder mathematically and gets you into 2nd year university level math.   Motion planning is a bit harder, and the AI to generate the motion planning target is not yet possible.

I think that is what makes robots interesting is that you can work at any level.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAArUYGtrfNLSZDRHVT08Fz%3D6eMdJDuuxCeOUwB_zeo8R1knf7Q%40mail.gmail.com.

[Stephen Williams]

On 6/13/26 10:17 AM, Chris Albertson wrote:

On Jun 13, 2026, at 7:45 AM, Ken Gregson <ken.g...@gmail.com> wrote:

Wayne,

Thanks, I'm familiar with some of that work.

I think we are at an inflection point in the availability of more generally useful robots and their accessibility to hobbyist and DIY developers. 

We can build nice puppets that must be teleoperated, and we can build what are basically CNC machine tools that run scripts for prerecorded motions.   Look at Elon Musk's robot demo at Universal Studios a while back.     The guy has nearly unlimited funds and an army of robotics engineers, and he literally had to bolt the robot’s feet to the floor so that robot dancers did not fall over.   Yes, fixed-feet screwed to the floor.      His other robots could walk on flat surfaces but not step up or down a curb and still needed full remote control.     And THAT is with a billion-dollar budget.

...

I don’t think we are very far along; we can make animated mannequins that are as smart as a toaster oven.   I’m not saying that as a criticism, but rather a challenge. The future is wide open, and one smart guy can still make a revolutionary discovery.   AI researchers in 2026 are like physics was before Isaac Newton.   The basics are still to be discovered.   It is good to work in this field.

We need good hardware widely available that is very inexpensive.  And to be working on the AI/ML/etc. in parallel as that hardware develops.

I'm working on the hardware side.  I can't wait to graduate to focusing on the software.

Stephen

But robots do make a good subject in a classroom.  Students need to learn how to solve problems like “How many M4 screws, in single shear,are needed to transmit 700 cm kg torque using a 75mm bolt circle?”   Or what rotation sensor accuracy is needed to give 1mm precision at the end effector or 100 other simple robotics engineering problems.   Those are both good high school level problems covering static cases.        Movement quickly gets harder mathematically and gets you into 2nd year university level math.   Motion planning is a bit harder, and the AI to generate the motion planning target is not yet possible.

I think that is what makes robots interesting is that you can work at any level.

--

Stephen D. Williams Founder: VolksDroid, Blue Scholar Foundation 650-450-8649 | fax:703-995-0407 | s...@lg.net | https://VolksDroid.org | https://BlueScholar.org | https://sdw.st/in

[Steve " 'dillo" Okay]

On Saturday, June 13, 2026 at 12:17:36 AM UTC-7 Tony Pratkanis wrote:

Hello All,

I have been reading up on papers on low-cost full scale mobile manipulation. There's a lot of smaller experiments such as the LeRobot and the LeKiwi, but these don't have a lot of range or height they can grasp at. I found several interesting papers.

XLeRobot - this is a mobile base by placing a Lekiwi base under an IKEA cart, then mounting LeRobot arms on top of the cart. See https://github.com/Vector-Wangel/XLeRobot. I think this is an interesting idea, but it has low payload (400 g per arm).

I have a stalled project that is using a Hackerbot base, some 8020 and one of these:
https://www.instacart.com/products/24770773-polder-undersink-storage-caddy-grey-17-x-10-1-2-x-16-1-2-each

as a similar system. 

Stalled only because I've either had work or work-related things to attend to. 

I like the idea of the dual arms, but really, I just need a mobile base with a tool-caddy to follow me around and hand me things when I need them. 

I've got a set of LeRobot arms, the base, the cameras, a scattering of RPi's . 
I just finished a short-but-intense 3-week contract(which involved LLaMa & VLA/VLM training & testing)  

Maybe I'll have time to get back to it for a bit now(?)  

'dillo

[Steve " 'dillo" Okay]

On Saturday, June 13, 2026 at 6:34:45 AM UTC-7 Ken Gregson wrote:

In the LeRobot "line" there's also the AlohaMini

https://github.com/liyiteng/AlohaMini. It uses a SO-ARM track axis for vertical linear motion/lift

https://github.com/avenhaus/SO-ARM100-Track-Axis

I'm building a couple variants of it AlohaMini "LeDomo" to keep my LeKiwi "Midori" company. Open question on how robust a PLA gear and toothed rack will be.

There are a growing number of "Tough" PLA blends out there, along w/ some other interesting materials like PCTG(not PETG) that are ballpark-ish cost to traditional PLA but

are significantly stronger. My point being: The "PLA" available in 2026 is not the PLA of 10 or even 5 years ago. 

'dillo

[Chris Albertson]

What you need to do is print some stuff then break it.

I was also curious and set up. motor and tsome gears and a dummy load made with a string wrapped around a drup with a weight in the drum.    I used a stepper motor to raise and lower the weight.

Firt off are the obvious results you wouild =knpow going in.

1. 3D prints are not are precision made as metal gears so you hhave to price a minimum gear tooth size such that the defacts are vert small compared to the size of the geat tooth.   You end up with a fairly large minimum size

2. Plastic is not nerly are strong as metal.  Lets say it is 5 times less strong.  So for the same strength you need to mke the plastic geat 5X larger thn a metal gear.   #1 and #2 actaul work toto gather, the ger need to be 5 or 6 times larger to large teeth “work”.

Next is a fast  that should have been obvious.  Gears have friction when they mesh, friction produces heat and this is the main caise of failure.   If you heat PLA even a tiny bit it becomes soft and then is VERY weak and deform and then you have run-awayf riction heating and soon teeth shear off

So you REALY, Absolutly want a kind of plastic that can take heat.  At the very least use ABS.      I think the effect is at the micro scale.  The entire gear did not heat up.  I think only at the small area what contact is made.

Het resistance plastic helped a whole bunch

“Eyeball engineering” does not work.  Our experiance with metal gear makes us think “these gear look rel stong” but we are wrong.  Do the math and yes they will be “fricking huge” and then use white grease and heat resistance plastic and they work well.

When the teeth are problely size as above, the failur pint is ALWAYS the hubs wher the connect to the motor shaft.  You wil strip out D-holes, splines and grub screws.   The only solution is to make metal hubs and press-fit them into the plastic gears.  You can make a metal hub by slicing a 3/4” metal rod (mild steel is hbest) drilling a center hoe and dilling an tapping from grub screws and th force it into a platic hole on the get, maybe use CA glue.     I’ve been using these for a couple years now.

One more thing:   Plastic timing belt pulleys seem to work MUCH better then plastic gears.   

Bittom line, PLA fails quickly.  The first thing is “spalling” and then shear off the teeth.   But of couse you can scale it up and go to Module 6 teach and 25mm face widths but then it can’t fit inside the robot.   Metal is MUCH more space effecent.

I built a CNC machine by. modifying a Harbor Freight manual mill.    The harder part was the vertical rack to move the milling head up and down.   The best solution was a ball screw.  They are nearly frictionless, don’t cost so much and can move hundreds of pound of weights.      SOmething like this might work. This is NOT a lead screw.  The nut is made of balls the recirculate.   The ball screw suported the vertical load but the two rails and four slider blocks contrain the motion to only one vertical.    You might even use four rails wit ht eball screw in the center.     You are basically building an elevator. that rides on ball bering.     You can easy get 1/100th inch accuracy wit this Amazon parts and a good closed-loop stepper motor.   I picked the sizes at random because I don’t know what you robot weighs or how far it needs to move.

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/7d8803b6-5ea3-4c79-a0c9-4bd6a1f553c5n%40googlegroups.com.

[Ken Gregson]

"There are a growing number of "Tough" PLA blends out there, along w/ some other interesting materials like PCTG(not PETG) that are ballpark-ish cost to traditional PLA but

are significantly stronger. My point being: The "PLA" available in 2026 is not the PLA of 10 or even 5 years ago. 

'dillo"

Good point, haven't experimented with PCTG yet. I've been pleased so far (except maybe with picking all that support material from the teeth in the track ;-). But -just in case- have printed a version with a flat instead of the track to which I could imagine affixing any number of things including a more robust CNC milled toothed rack. It's also perfect for working on the base without the "distraction" of the arms.

So many possibilities!
- Ken

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/7d8803b6-5ea3-4c79-a0c9-4bd6a1f553c5n%40googlegroups.com.

[Chris Albertson]

If you are talking about gears, heat tolerance is the primary consideration.    Strength is not nearly as important because if it needs to be stronger, just make it bigger with a wider face width.   

About a rack, just use a lead screw.   A rack used to lift means you are constantly needing to supply torque,  Lead screws are not back drivable

Belt drive work better then gears if you are printing, they last forever and have the same torque and power requirements as racks.  Bit they are back drivable so the load falls if power is lost.    

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAArUYGvjNNfew959tb77f1CaGtJdtMVbOM_7ef-bKMWXTZ%3D7Jw%40mail.gmail.com.

[Tony Pratkanis]

Hello All,

Found another one: https://vulcanrobotics.ai/. This is a mecanum base + feetech servo arms. Appears open source.

Regards,

Tony

[Chris Albertson]

The mecanum base is very good if you have a hardwood floor, not so otherwise.     The most robust and reliable of the holonomic bases is four steerable wheels. 

Watch out for using servos for hands.    While they seem very intuitive and simple, they have a HUGE flaw: Servos cannot control force.     Think of this example:   A full unopened can of soda and an empty can of soda, and the robot is holding one in each hand.      The finger positions are exactly the same in each hand, but the full can needs more pressure to keep it from slipping; if the same pressure were applied to the empty can, the can would be crushed.

You can sort of fake it by using rubber skin or rubber in the joints and the can we can control pressure with position  .  Springs convert distance to force by Hook’s Law.   But this is crude.   

You really do want one of those newfangled fancy motors where you can specify force, position, and velocity all at once.    This kind of controller is available open source if you like.

If you are experimenting, I think the first step is to bolt the arm to a tabletop.  Just one arm and no mobile base.

I see the controller for humanoid robots as having three layers (1) the first layer is for balance and walking and is created using RL (reinforcement learning) and (2) the second layer is a trained library of maybe 100+ tasks like picking up objects or tipping water out of a container, turning a screwdriver, and so on.     These are likely trained by imitation learning with teleoperation and (3) the top level is the planner that finds a sequence of tasks to will solve some kind of goal.   This is a higher level AI.

I’d want so well-defined interfaces between the layers.

On Jul 26, 2026, at 6:53 PM, Tony Pratkanis <a...@pratkanis.co> wrote:

Hello All,

Found another one: https://vulcanrobotics.ai/. This is a mecanum base + feetech servo arms. Appears open source.

Regards,

Tony

On Saturday, June 13th, 2026 at 12:17 AM, Tony Pratkanis <a...@pratkanis.co> wrote:

Hello All,

I have been reading up on papers on low-cost full scale mobile manipulation. There's a lot of smaller experiments such as the LeRobot and the LeKiwi, but these don't have a lot of range or height they can grasp at. I found several interesting papers.

XLeRobot - this is a mobile base by placing a Lekiwi base under an IKEA cart, then mounting LeRobot arms on top of the cart. See https://github.com/Vector-Wangel/XLeRobot. I think this is an interesting idea, but it has low payload (400 g per arm).

AhaRobot - this is move complex to build but it features elevators and a SCARA-like arm design. See: https://arxiv.org/abs/2503.10070. This layout of SCARA + elevators means that joints are not under high torque at the base of the arms. It also has a higher payload (1.5 kg per arm).

There's also Nori Bot which appears to be a hybrid of these two approaches. https://arxiv.org/html/2605.16537v1. Also payload of 400 g per arm.

For payload, I have looked into average objects that people carry on a daily basis. One of the heaviest objects appears to be a milk jug at around 4-6 kg, so I think 6 kg total payload would be a good goal to achieve.

Anyway, has anyone else seen or built anything similar to one of these?

Thanks,

Tony

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/7FpsKicJdqJ1E-gtdpqwcO2OfLEGSyFknEvJ0Ys7pMgGAjQWHilz9mo33z6hwcgl0CMwhXJE_mPdsmcua7IDGTOOYRX9rJu4ZRPKXYCLCMQ%3D%40pratkanis.co.

[Tony Pratkanis]

Hair and dirt is another issue at scale with mecanum and omniwheels getting stuck in the wheels.

As for the servos, the Feetech and other Dynamixel clones have the ability to set the max force that is adjustable in real time as a command over the serial bus. The gotcha is that there is a lot of blacklash in the gearing so it is not the kind of precise control you get with a QDD system. The noribot paper talks about how to get precise force control with rubber grippers dampening out the backlash.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/E8232CB2-F02B-4690-BC52-93192A0F04E6%40gmail.com.

[Dave Everett]

On Mon, 27 Jul 2026 at 12:39, Chris Albertson <alberts...@gmail.com> wrote:

Watch out for using servos for hands.    While they seem very intuitive and simple, they have a HUGE flaw: Servos cannot control force.     Think of this example:   A full unopened can of soda and an empty can of soda, and the robot is holding one in each hand.      The finger positions are exactly the same in each hand, but the full can needs more pressure to keep it from slipping; if the same pressure were applied to the empty can, the can would be crushed.

Most of my work over the last year has been collecting full and empty soft drink cans. My gripper all throughout this period has been servo driven, I don't even have pressure sensors, just limit switches that trigger as the finger plates compress. I used a very soft neoprene on the these plates for gripping. I have found the same force for full or empty cans is fine. I intend to add pressure sensing later, and a complete redesign of the gripper, but my focus has been on identifying, locating in space and then picking up at the moment.

It's worth noting that the robot will still have no way to know what force to apply, unless it is able to determine the weight of the can. There are obviously ways to handle that. In the first version, built about 15 years ago, the wrist had a pivot that was sprung and would operate a limit switch if the can was over a certain weight, indicating it was not empty. The robot's only job was to go to a common area at the university, locate cans, check if they were empty and if so, take them to the recycling bin in another part of the uni.

Dave

[Alan Timm]

Take a look at xlerobot.

https://xlerobot.readthedocs.io/en/latest/index.html

And they've even assembled kits for it now.

https://shop.wowrobo.com/products/xlerobot-dual-arm-mobile-household-robot-kit?variant=47297659961561

The lowest cost entry I can find.

On Saturday, June 13, 2026 at 12:17:36 AM UTC-7 Tony Pratkanis wrote:

[Chris Albertson]

The more specialized the task, the simpler the robot arm can be.     I think for a realistic hobbyist, you have to say you are not ever in your lifetime going to build a “plumber”, that is a robot who shows up at a house, assesses the problem, then gets tools and materials and fixes the issue all without any supervision.    

A more realistic task that might take only a year or two is a pick-and-place robot arm where you tell it “pick up the fork and place it on the plate”.   The real task you would do is not building the arm because you can buy those at any price point you want.  The task is implementing imitation learning and  reasoning about physical objects.

If the goal is to have a trained arm, start with the simplest and cheapest one you can find and just one of them.   Your first tasks will be very simple, so a simple arm will work.

“Blocks World” has been an interesting AI/Robotics playground for decades.   You tell the robot to stack blocks like “stack three blocks so the blue one is in the middle”.    I’d argue that if you can not make that work, owning a more complex two-armed mobile robot is pointless.

On Jul 27, 2026, at 10:02 AM, Alan Timm <gest...@gmail.com> wrote:

Take a look at xlerobot.

https://xlerobot.readthedocs.io/en/latest/index.html

And they've even assembled kits for it now.

https://shop.wowrobo.com/products/xlerobot-dual-arm-mobile-household-robot-kit?variant=47297659961561

The lowest cost entry I can find.

<screenshot_20260727_100137.jpg>

On Saturday, June 13, 2026 at 12:17:36 AM UTC-7 Tony Pratkanis wrote:

Hello All,

I have been reading up on papers on low-cost full scale mobile manipulation. There's a lot of smaller experiments such as the LeRobot and the LeKiwi, but these don't have a lot of range or height they can grasp at. I found several interesting papers.

XLeRobot - this is a mobile base by placing a Lekiwi base under an IKEA cart, then mounting LeRobot arms on top of the cart. See https://github.com/Vector-Wangel/XLeRobot. I think this is an interesting idea, but it has low payload (400 g per arm).

AhaRobot - this is move complex to build but it features elevators and a SCARA-like arm design. See: https://arxiv.org/abs/2503.10070. This layout of SCARA + elevators means that joints are not under high torque at the base of the arms. It also has a higher payload (1.5 kg per arm).

There's also Nori Bot which appears to be a hybrid of these two approaches. https://arxiv.org/html/2605.16537v1. Also payload of 400 g per arm.

For payload, I have looked into average objects that people carry on a daily basis. One of the heaviest objects appears to be a milk jug at around 4-6 kg, so I think 6 kg total payload would be a good goal to achieve.

Anyway, has anyone else seen or built anything similar to one of these?

Thanks,

Tony

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/6588d762-3478-4b87-b90e-b0d60f50d3ecn%40googlegroups.com.
<screenshot_20260727_100137.jpg>

[Ken Gregson]

There's also the AlohaMini to consider

https://github.com/liyiteng/AlohaMini

Similar to XLeRonot with a z-axis lift for the arms. 

I'm working on one now.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/C4AC8390-EEF1-4222-B62A-8C1ABF18F804%40gmail.com.

[Dave Everett]

On Tue, 28 Jul 2026 at 04:07, Ken Gregson <ken.g...@gmail.com> wrote:

There's also the AlohaMini to consider

https://github.com/liyiteng/AlohaMini

Similar to XLeRonot with a z-axis lift for the arms. 

Their BOM is WAY off the mark, just the servos are about double the cost listed, I didn't check further because I was shocked by how cheap they were claiming they were.

Dave

[Chris Albertson]

On Jul 27, 2026, at 6:53 PM, Dave Everett <daveev...@gmail.com> wrote:

Their BOM is WAY off the mark, just the servos are about double the cost listed, I didn't check further because I was shocked by how cheap they were claiming they were.

Are we looking at the same BOM?  I noticed there is a version 1 and 2 of the robot and I checked out V2 and the BOM seems to be spot on.          I would not build the entire robot, try just one arm and see how it goes.   It is good that they included all the .STP files so you could modify the parts

This is what I was looking at:

[Dave Everett]

'

Yes that's what I was looking at. Under the Mobile Base section, they list the STS-3095 servo as about $50. I followed the link and the price was $72.64 and you must buy a minimum of 3 servos. 

Looking down into the ARM section, the ST-3215 servos are listed as $50.37, they require qty 6, so at least that is withing the requirement of the seller, but that is still $72.64 each, not $50.

The ST-3215 servos are listed as $19, but the seller page shows $23.03, that is about $1 less than I am currently paying, I'd just ordered 6 for a projects. However when selecting, the C0018 version specified is not available, that's not the developers fault, but it is an issue. They also charge shipping if the shipping cost goes over $30, not certain if that would occur as I was not able to enter any servos into the cart.

The BOM was made on the 6th of June, I'd go through more items, but I suspect the cost will probably be considerably higher than listed.

I know these things can change, but that is a substantial increase in a month. I've just gone through the process of building a BOM for a project and I understand it is annoying, so I appreciate the work that has gone in to this. 

Dave

[Ken Gregson]

AlohaMini1 has same arms, servos as the XLeRobot (+1 for the lift) - should be relatively the same? Haven't priced out the AlohaMini2 with more capable SO-ARM-200 and base.

Last order I made from WowRobo Robotics for: 

Feetech STS3215 C018 Servo – 12V 30KG High Torque Servo for SO-ARM100/101 × 3

Priced out to $13.99 a piece (in April). Shipping costs can be steep for small quantities but only add a couple bucks per on larger orders. Tariffs have been all over the place.

Inflation and memory premiums are also things. Wasn't that long ago you could get a RPi5 16G for $120 (now $300) or a Jetson Orin Nano 8G for $250 (now $400). Thanks AI!

- Ken

--

You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAG61pbfH-K6zRwmcfMXab4aKk%2BrmSNPkpyvZ1X5aX4hmbSoSmg%40mail.gmail.com.

[Dave Everett]

On Tue, 28 Jul 2026 at 17:05, Ken Gregson <ken.g...@gmail.com> wrote:

Last order I made from WowRobo Robotics for: 

Feetech STS3215 C018 Servo – 12V 30KG High Torque Servo for SO-ARM100/101 × 3

Priced out to $13.99 a piece (in April). Shipping costs can be steep for small quantities but only add a couple bucks per on larger orders. Tariffs have been all over the place.

Those prices are amazing. I just ordered a couple of the 50kg-cm servos. With shipping though it came to just under $200 AUD. 

Dave

[Chris Albertson]

This is really odd.  I click the “buy” link, and I see the $50 servo for about $50.      The same link is showing different prices to each of us???    I live in the US, California.  The site figured this out and shows in US Dollars.

But if you really want to build this, you could use whatever servos you like, get some from Amazon, or use a different motor.   The design is open source, so you can change the way the servos mount and use completely different ones.   One problem was designs like this. 

But as said, I don’t see why anyone would build this complete robot. Just do one arm first and attach it to the table using clamps or screws, no mobile base.  Getting one arm trained and working will take more than a year, and you will gain enough experience to know what modifications you want to add to the second arm.

BTW, I’ve been looking at V-JEPA models, but just running the open source model needs a 4090 GPU.   I’m going to look at smaller models and also into VL-JEPA.    But this is the future, I’m interested in walking, and for that I only need a “tiny” model that is RL-trained.

But if you build an arm first, you are jumping off into  “huge model territory" from day one, unless all you want is teleoperation.

Yes that's what I was looking at. Under the Mobile Base section, they list the STS-3095 servo as about $50. I followed the link and the price was $72.64 and you must buy a minimum of 3 servos. 

Looking down into the ARM section, the ST-3215 servos are listed as $50.37, they require qty 6, so at least that is withing the requirement of the seller, but that is still $72.64 each, not $50.

The ST-3215 servos are listed as $19, but the seller page shows $23.03, that is about $1 less than I am currently paying, I'd just ordered 6 for a projects. However when selecting, the C0018 version specified is not available, that's not the developers fault, but it is an issue. They also charge shipping if the shipping cost goes over $30, not certain if that would occur as I was not able to enter any servos into the cart.

The BOM was made on the 6th of June, I'd go through more items, but I suspect the cost will probably be considerably higher than listed.

I know these things can change, but that is a substantial increase in a month. I've just gone through the process of building a BOM for a project and I understand it is annoying, so I appreciate the work that has gone in to this. 

Dave

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAG61pbf4P7mNhaZsg0umZR8CnOBGz2SyxWFgFW%2BmYD5SsRkaag%40mail.gmail.com.

[Alan Timm]

Since we're talking about low(er) cost serial bus servo options.

A new HiWonder NexArm crossed my feed today.  

Nothing really special about it compared to The Waveshare or Aloha min arms except...

https://www.hiwonder.com/products/nexarm?variant=42737911201879

Looks like they have a 65kg/cm serial bus servos for $50.

https://www.hiwonder.com/products/hx-65hm?variant=42066913132631&_pos=1&_sid=efdbf7f0a&_ss=r

https://drive.google.com/drive/folders/1yaZ8iRYgWncdHPopioo7OYuHoGhrMxCi

They're a little on the wide side, but that's half the price of the feetech 60kg servos that were on my radar.

Alan

[Chris Albertson]

What exactly are you building?  It is not hard to bolt 6 ot 7 servos together and make alarm.      But the precision is the sum of all the errors in the chain.  The weak link is all the crappy veriable resisters ued to measure rotation angle.   Then after that are the crappy bearings used in cheaper servos

Arm are call “serial link” for a reason.  Each motoris coneted to a mechanical link then the the next motor attaches to that and so on for 6 or 7 motors.  Ther wobble in one ands to the nest and so on.     You end of with a lot of “play” byu the time you get to the end.      My robot-dog only has three links per lleg and after an hour the bearing get looser and ther is a few mm of play at the feet.

But this ois OK as a revision zero test and it lets you verify your software.    Then build a better one.      Eventually you find you need to spend $100+ in each motorand that the cheapest sensor are not good enough and you need better bearing.      But you never know untill after that first cheap build

Another way to go is to buy a generai analog servo and remove the little controller PCB and run the motor leased to a normal motor controller.  Now you have a high-ratio gear motor and then use a low-cost magnetic rotation sensor on the joint.. Now you have a high precision joint

I stink small stepper motoer work better in the shoulder where the arm does not need to cary the weight.  But steppers are too heavy for other parts of ther arm

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/ec9536d7-2b36-4700-9fba-8c0fba622d3dn%40googlegroups.com.

[Tony Pratkanis]

If you actually want high precision you have two options:

1. Develop a high precision 6-DOF sensing technology. Optical tracking works okay. Then build a control loop on top of the end effector pose. As long as the relative precision is ok absolute is achievable.
2. Have a high precision robot arm. Probably the cheapest is openarm.dev or one of the rebot or trossen arms.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/03FC5E5E-4B32-42FC-B9FE-8C02EC582378%40gmail.com.

[Chris Albertson]

The problem with servos, even if you have perfect optical tracking, is that you can only command them to within a degree or so because their internal rotation sensor is so poor.  It is literally just a metal “brush” riding on a carbon resistor.  It is a low-precision analog device.   Then there is the issue of bearing play and large amounts gear backlash

So with your good optical tracking you are close and the software command to move 2mm the arm does nothing, so you say 2mm more and now it jumps 6 mm.

You would think you could home in, but the whole system has noise.  It is good enough for many uses but not for precision work.

Yes, buying an arm is the easiest way.   But just like those $5,000 humanoid robots, what do you do with it after you buy it?  You will spend years trying to get it to do a useful task.   Why not work on that first?  No money is needed; then, after you have it working, buy some hardware.

[Dave Everett]

IOn Jul 28, 2026, at 9:43 AM, Alan Timm <gest...@gmail.com> wrote:

Since we're talking about low(er) cost serial bus servo options.

A new HiWonder NexArm crossed my feed today.  

Nothing really special about it compared to The Waveshare or Aloha min arms except...

https://www.hiwonder.com/products/nexarm?variant=42737911201879

I've asked them the lifting capacity at full reach. Apparently that was not worth putting in the specification, but the name of the arm was.

Dave

[Dave Everett]

On Wed, 29 Jul 2026 at 07:17, Chris Albertson <alberts...@gmail.com> wrote:

So with your good optical tracking you are close and the software command to move 2mm the arm does nothing, so you say 2mm more and now it jumps 6 mm.

You would think you could home in, but the whole system has noise.  It is good enough for many uses but not for precision work.

That's why you run a closed-loop, recomputing the desired position as you go based on sensor data. Nothing at our level is particularly precise. Sensors vary, vision systems jump around, distance calculations are affected, servos have finite repeatability, that's just how it is for what the average hobbyist can afford. In general this is fine. I'm not building a robot to insert pencil leads into the tube, I want to pick up a can that is 20mm smaller than the open gripper width. 

Yes, buying an arm is the easiest way.   But just like those $5,000 humanoid robots, what do you do with it after you buy it?  You will spend years trying to get it to do a useful task.   Why not work on that first?  No money is needed; then, after you have it working, buy some hardware.

I'm missing something here, how do you work on it without hardware? Are you talking about simulation?

 

Dave

[Alan Timm]

That's exactly why I love these serial bus servos.  The cheapest ones do still use a potentiometer, but alot of the feetch and hiwonder ones use a diametric magnetic encoder.

[Chris Albertson]

I'm missing something here, how do you work on it without hardware? Are you talking about simulation?

Yes that is pretty much how it is done today.  Especially after the move to training rather than hand-coding movements.  You need a simulation so you can let it run on thousands or even millions of trials.  Then later move to real hardware

The older method using inverse kinematics can work too but typically solutions are brittle.  A lot depends on how complex the problem is and if you want the robot to handle “everything” or only a few well-defined cases.

 

[Dave Everett]

On Wed, 29 Jul 2026 at 10:07, Chris Albertson <alberts...@gmail.com> wrote:

I'm missing something here, how do you work on it without hardware? Are you talking about simulation?

Yes that is pretty much how it is done today.  

I've seen many simulations of millions of runs fail on real hardware. I'm still not convinced it is the way to go. For example, a friend has built one of those Disney ducklings that uses feetech 19kg-cm servos. He used code developed from an Isaac sim. The sim works great, the real hardware, not at all.

Dave

[Chris Albertson]

For a robot arm, you could do it either way.  The traditional way is to use a depth camera to identify the object to be grasped look up a plan for that kind of object, and then find the point in 3D space and orientation to place the arm and then use inverse kinematics to calculate the joint angles and find a path in joint-space to move the arm while checking for collisions with self or the environment.  This method is well-known.

But for robot legs and walking, I did that using the above hard-coded method for a quadruped, but  it gets too hard if you want to do things like walk outdoors on uneven dirt paths.  Training results in a more robust solution.

As for simulations not working, Isaac is very widely used all over the world in tens of thousands of places. It basically works for all those people.   If a sim to real project fails, the #1 most likely problem is an oversimplified physics model that does not model contact dynamics and friction, random perturbation and structure deformation under load, sensor noise, and a dozen other things.

Maybe you can say the model was overfitting and did not generalize with a simplified physics model.

 

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAG61pbcb845W1E8LJVoacjndiEdYW7E9%2BEetpMojFV9VBTuODg%40mail.gmail.com.

[Dave Everett]

On Wed, 29 Jul 2026 at 15:24, Chris Albertson <alberts...@gmail.com> wrote:

For a robot arm, you could do it either way.  The traditional way is to use a depth camera to identify the object to be grasped look up a plan for that kind of object, and then find the point in 3D space and orientation to place the arm and then use inverse kinematics to calculate the joint angles and find a path in joint-space to move the arm while checking for collisions with self or the environment.  This method is well-known.

That's pretty much what I'm doing, a kinect to do depth and for the tensorflow image to a coral accelerator. The position of the can is then used to servo the camera neck and the body rotate to put it inline with the arm axis.

As for simulations not working, Isaac is very widely used all over the world in tens of thousands of places. It basically works for all those people.

Except for all the people that it doesn't work for, like my friend and also Skyentific and others. 

 

If a sim to real project fails, the #1 most likely problem is an oversimplified physics model that does not model contact dynamics and friction, random perturbation and structure deformation under load, sensor noise, and a dozen other things.

Yes that has always been the case and still remains an issue as it is very difficult to ensure all the physics of your robot are right. That's why I prefer real world, the physics are correct, and you will find that out very quickly when things go wrong :)

Dave

[Dave Everett]

Apparently the lift capacity is 500g.

[Tony Pratkanis]

Yes for Feetech 30 kg cm-based arms that is the usual limit. E.g. for the SO101.

--

You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAG61pbcZXk6i7ng5HmQq%2BwPNDWkSx1RvxOoB7So1tSAw-edHSw%40mail.gmail.com.

[Stephen Williams]

Nice references, approaches, and parts references.  I am going in a mostly different direction, which is challenging but fun and may be rewarding.

I just can't get interested in using cheap servos, and I don't want to pay for expensive servos.  I have some of each, and will use them in places for some prototypes or toy projects.  But for interesting-to-me robotic drive mechanisms, I am concentrating on end-to-end closed loop control of the drive mechanism.  This is equivalent to a servo, just not self-contained, but also not as constrained.

I now have a new take on my now-old SPIVT approach.  If this works as well as I expect, it should get me where I have been trying to go for a long time: drastically lowered weight, cost, and complexity with a quiet drive system.

The point of the SPIVT was to create a RIVT, a reversible infinitely variable transmission, which can then modulate an input rotation to servo-like output.  That is interesting because you can then drive multiple joints / degrees of freedom off of a shared motor.  Fewer motors, fewer power controllers, fewer heavy noisy gear reducers which is less weight, less cost, etc.  My first complete approach was a bit bulky, had too many parts.  My new approach will fix that, if the complete design works well.

A number of people have encouraged me to finish my hand design, and I will soon.  But this drive mechanism is more exciting.  It is exactly right for driving the hand for one thing.  I will see how much I can miniaturize & bundle drive units.  It will scale up easily.  I'd like to produce tiny versions too.

I have most of the needed maker tools now, including a new Rotatrix trackball for CAD, sheet metal cutting & welding, metal & plastic CNC, and a continuous CF FDM printer.  I have evolved the solution for a little while, so now it is on to design & prototyping.

https://rotatrix.com/

Then I will be back on the hand.

Stephen

On 7/26/26 6:53 PM, Tony Pratkanis wrote:

Hello All,

Found another one: https://vulcanrobotics.ai/. This is a mecanum base + feetech servo arms. Appears open source.

Regards,

Tony

On Saturday, June 13th, 2026 at 12:17 AM, Tony Pratkanis <a...@pratkanis.co> wrote:

Hello All,

I have been reading up on papers on low-cost full scale mobile manipulation. There's a lot of smaller experiments such as the LeRobot and the LeKiwi, but these don't have a lot of range or height they can grasp at. I found several interesting papers.

XLeRobot - this is a mobile base by placing a Lekiwi base under an IKEA cart, then mounting LeRobot arms on top of the cart. See https://github.com/Vector-Wangel/XLeRobot. I think this is an interesting idea, but it has low payload (400 g per arm).

AhaRobot - this is move complex to build but it features elevators and a SCARA-like arm design. See: https://arxiv.org/abs/2503.10070. This layout of SCARA + elevators means that joints are not under high torque at the base of the arms. It also has a higher payload (1.5 kg per arm).

There's also Nori Bot which appears to be a hybrid of these two approaches. https://arxiv.org/html/2605.16537v1. Also payload of 400 g per arm.

For payload, I have looked into average objects that people carry on a daily basis. One of the heaviest objects appears to be a milk jug at around 4-6 kg, so I think 6 kg total payload would be a good goal to achieve.

Anyway, has anyone else seen or built anything similar to one of these?

Thanks,

Tony

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/7FpsKicJdqJ1E-gtdpqwcO2OfLEGSyFknEvJ0Ys7pMgGAjQWHilz9mo33z6hwcgl0CMwhXJE_mPdsmcua7IDGTOOYRX9rJu4ZRPKXYCLCMQ%3D%40pratkanis.co.

--

Stephen D. Williams Founder: VolksDroid, Blue Scholar Foundation 650-450-8649 | fax:703-995-0407 | s...@lg.net | https://VolksDroid.org | https://BlueScholar.org | https://sdw.st/in

[Chris Albertson]

I have just started learning how to record joint torques and speeds.    This is about locomotion with legs.   

There are torque swings going from positive to negative direction with no or only little rotation.   I’m wondering how a variable-speed transmission can send wildly varying and different torques to multiple joints using only one motor.

I’m also wondering how backdrivability can work.  Maybe the CVT has some programmable variable slip. But then you’d eat power with friction.   Veriable speed seems od because robots either arms and legs mostly run near zero speed.   Carrying a box in the arms is a good example,   Force is needed at zero RPM.

The other problem I have is the force required.    I’m looking at a full-size humanoid with normal human weight and see torque spikes going up to a few hundred Newton-meters.  A single motor would have to be large enough for the worst case where all joints need peak torque simultaneously.

I don’t want to say “you can’t do this because I personally can’t figure out how”.     Can you give us a hint about how this works, a schematic sketch maybe?

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/a147c061-99fe-4bf9-96d1-a1afc0b0aa9b%40lig.net.

[Stephen Williams]

This is the SPIVT patent.  I really fell short by not having 'Reversible' in the name.  This makes most of the points, although this particular mechanism is clunky, too big, etc.  I set it aside while I worked on other things, but now I'm back on it with my new take.  My new idea is related to this, more or less a derivative, but much better for these purposes.  But I'm not going to talk about it in detail until I prototype, file a new patent when I get it working well.  It is a straightforward idea.

https://patents.google.com/patent/US20190128390A1

This is friction based.  There are only a few ways (about 19 now I think, several of them from one French guy) to make an IVT.  The simpler approaches are friction based.  There are almost no ways to create an RIVT - Reversible infinitely variable transmission.  The prominent one that everyone has seen in action is using hydraulics with swash plate motor + transmissions, such as in a tow truck bed & winch control.  Not very attractive for robotics.  The SPIVT claims the general solution of RIVT = Input * CVT + -Input / 2.  The example SPIVT mechanism is not perfect.  My new approach is better, in a certain way.

There will be some power loss from friction, but this can be optimized.  My intuitive sense of the possible efficiency of that is from decades of high-speed long-distance inline skating.  Polyurethane of a well-matched durometer (90a for instance) is amazing: grippy, but doesn't deform much, and lasts for hundreds of miles on rough pavement or concrete with 200+lbs of force.  Rolling resistance from deformation uses power, but with the right combination, it is fairly minimal.  Polymers on aluminum (or magnesium I think) is well-known to be a long-wearing combination, far more forgiving than roadways.

Variable speed / torque is exactly what is needed for robotic limbs operating at near zero speeds.  At low speeds, the torque is very high so that it is relatively easy to reach & hold a fine-tuned position.

Total force needs to be available, but you can reach that in a number of ways.  One method I included in that patent is to have multiple motor outputs mechanically routed to power joints that need maximum power for a moment.  In most robots, you'd want motors for each region, so perhaps one for wrist + hand, one or more for shoulder + elbow, etc.  A detailed face uses very little torque, except for jaw, tongue, so a single motor could run all of those degrees of freedom.

Also, torque can be stored, one way or another, to assist with peak demands.  This could be a net efficiency & weight benefit.

One related problem with driving multiple joints with a single motor is distributing the power.  If you place a number of transmissions together, you can just use belts, gears, or similar.  A typical output is just a set of tendons (string, cable, belt) which can be routed, perhaps through Bowden Cables.  Routing an infinitely rotating power source, such as a derivative of the motor output, to distant SPIVT inputs, that is more difficult.  Past approaches involve belts with sheaves (pulleys) redirecting power at pivot points (like the center of a shoulder or elbow joint), hydraulics, or speedometer cable like flexible rotating shafts.  All of these are annoying & difficult to deal with.

Yesterday, I conceived of a new solution for this that fits my patterns.  Simple, clean, flexible, very lightweight, cheap.  This will let me easily ship rotating power suitable for SPIVT input flexibly over any distance.  I even have a funny aspect of it that will be fun.  That has to go in the later bucket too though.  Sorry about that.  But it is coming!

I'm trying to protect the most interesting ideas from being coopted, stolen, etc. by big annoying entities.  Once I have that locked down, some or all of this should be easily available for hobbyist & startup use free or almost free.  Still working out the details, but I'm starting to get a pile of interesting elements.

If someone really really wants to know details and doesn't mind an NDA, and can reasonably critique, validate, etc., I'm open to that.

Stephen

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/0DE33185-5007-4B86-8226-593B0802351E%40gmail.com.