Tesla Optimus Sub-Prime Robot

[MC3OZ]

Discussion on Optimus Bot

There is another TMC thread here:- Optimus Bot Future

I'll reserve this first post for linking other relevant areas of TMC.

 

[MC3OZ]

Since there is some discussion on the form of the Bot, my thoughts are Humanoid+ is worth exploring.

By that I mean the bot has a basic human form and we dressed in the right skins looks more human.

Under the skin the Robot has attachment points for tool arms, vacuum cleaners, air compressors, etc.

Vacuum cleaners, air compressors and fluid dispersers would mount on the back of the bot.

Additional tool arms could plug into the side of the bot at various mounting points, shoulder, under-arm-waist, thigh.
Up to 4 mounting points on each side, but probably a maximum of 4 additional tool arms at one time, with the 2 regular arms merely holding stuff.

For example, this could mean the bot could torque 4 bolts at the same time while 2 arms hold an object in place.

Each tool arm can fit a number of tools:-

• Torque Wrenches
• Screw Drivers
• Drills
• Hammers
• Sanding
• Filing
• Cutting
• Soldering
• Welding (probably with additional kit)
• etc

Other tool arms for:-

• Vacuum cleaning
• Washing
• Polishing
• Dispensing fluid.
• Nail Gun

Tubes containing fluid/air can pass around the side of the bot into the tool arm.

The regular arms are used for plugging in tool arms and tools as needed.

For example a bot could wash windscreens and top up tyres at Superchargers, it could note the thread depth and any alignment issues and inform the customer about the state of the tyres.

 

[MC3OZ]

A version of the bot could also be used to monitor human health and do basic tasks at hospitals and aged care homes.

We want to monitor:--

• Blood pressure
• Heart rate
• Blood oxygen levels
• Temperature
• General observations - skin colour, eyes etc

We want this version to look more human and less industrial, it can still have gear mounted on it's back.

In this case the arms are probably custom arms for a medical bot, if a medical bot is redeployed to an industrial usage, we would need to remove the skin, and some of the kit, fit new arms - but under the skin industrial and medical are the same.

There is a while service category with bots assisting vision impaired people especially on overseas trips. We want this version for the bot to look a bit more human and blend in as far as possible. It needs no special equipment, just vision and sound should be sufficient.

The assistance bot is also a companion, we may or may not need a form that looks like a dog, but people might get used to the idea of a humanoid companion, and that humanoid companion checking their health from time-to-time.

 

[MC3OZ]

Additional thoughts:

The bot can charge-on-the-job their via wireless charging, being plugged into a A/C power source or via a visiting mobile fast charger,
So 24x7/365 operation is possible,

This increase the chance that for factory work, 3 kWh of cells is more than enough.

Where we might see "Bot Range Anxiety" is in outdoor mobile applications.
Battery cells are likely to be in the legs and lower body, any arms are likely to be light weight, so moving the arms is energy efficient.
But I wonder how efficient the bot will be when walking?
It will get better range, by having retractable wheels in it's legs, or by having a skateboard that it can step onto,

But many outdoor environments like the beach, snow, ice, waking trails, rock climbing, are likely to be no-go areas for sometime.

This reminded me of an earlier discussion when someone suggested an "assistance-vest" for blind people, Rather than having a guide dog/robot, the blind person simply wears the vest. For international travel, they can walk onto the plane wearing the vest, and have a spare in their carry on luggage.

To clear airport security the vest needs to be the right design, a thin lightweight battery and low energy consumption is ideal.
But clearing airport security with a robot or a robot dog, is probably an even harder task, and both would take up significant space in an aircraft cabin.

 

[jeewee3000]

I love how going down the rabbit hole of possible Optimus utility reads like a scifi novel.

I do however think Tesla wants the hardware to be more general at first, since they place a heavy emphasis on reducing the cost of manufacturing these bots. Therefore: the less versions of the item to be produced, the better.

Adding mounting points is something they should indeed do either way, and then the customer (or an aftermarket) can create tools to stack onto the bot.

This is however specialized work and I don't think many are capable. The idea behind Optimus is that our world is already tailored for humans, so if you create an average human sized bot with similar capabilities he can do anything.

Once you start creating specialized bots (for example screwdriver bot that can hold things in place with two hands and fasten multiple bolts at the same time) the question that needs answering is: is the usage of Optimus the most efficiënt way to perform the task in the production line? I'd say no, since those tasks are generally done better by KUKA robots that live to fasten bolts.

Optimus should be used for things that are currently very difficult for robots:
- changing environment (need of perception)
- complex navigation (need of FSD)
- changing tasks
...

What (useful) tasks it will do in the beginning, I have a hard time imagining. But in the long run: everything except running the country, basically.

Exciting times.

 

[jeewee3000]

This reminded me of an earlier discussion when someone suggested an "assistance-vest" for blind people, Rather than having a guide dog/robot, the blind person simply wears the vest. For international travel, they can walk onto the plane wearing the vest, and have a spare in their carry on luggage.

To clear airport security the vest needs to be the right design, a thin lightweight battery and low energy consumption is ideal.
But clearing airport security with a robot or a robot dog, is probably an even harder task, and both would take up significant space in an aircraft cabin.

Can you explain what this assistance-vest is/does? I don't understand how it's supposed to work.

 

[MC3OZ]

Can you explain what this assistance-vest is/does? I don't understand how it's supposed to work.

Uses cameras and neural nets to map the world, then informs the blind persion via an ear bud.

E.g Unknown object 3ft away on the left, door to room 8 on right, you mother just entered the room.

So it is essentially vision neural nets and it needs to understand enough of the world to allow the blind person to safely move around.

The blind person might use a cane as an additional aid.

Blind people can't easily take guide dogs on international trips.

The vest does a similar job to a robot.

Like FSD and the bot, solving vision is the key.

Estimating the distance to unknown objects is difficult for vision, but known objects may be a reference.

Mapping navigational space or when and where it is safe to cross the road is similar to FSD.

It needs to be very conservative about things like crossing the road.

 

[MC3OZ]

I love how going down the rabbit hole of possible Optimus utility reads like a scifi novel.

I do however think Tesla wants the hardware to be more general at first, since they place a heavy emphasis on reducing the cost of manufacturing these bots. Therefore: the less versions of the item to be produced, the better.

Adding mounting points is something they should indeed do either way, and then the customer (or an aftermarket) can create tools to stack onto the bot.

This is however specialized work and I don't think many are capable. The idea behind Optimus is that our world is already tailored for humans, so if you create an average human sized bot with similar capabilities he can do anything.

Once you start creating specialized bots (for example screwdriver bot that can hold things in place with two hands and fasten multiple bolts at the same time) the question that needs answering is: is the usage of Optimus the most efficiënt way to perform the task in the production line? I'd say no, since those tasks are generally done better by KUKA robots that live to fasten bolts.

Optimus should be used for things that are currently very difficult for robots:
- changing environment (need of perception)
- complex navigation (need of FSD)
- changing tasks
...

What (useful) tasks it will do in the beginning, I have a hard time imagining. But in the long run: everything except running the country, basically.

Exciting times.

Attachment points were a way of powering tools via the bot and a hedge against issues with the hand.

How well can the bot hand grasp tools? And how precise is the end result?

To be honest this is more of a group brainstorming process, we can just come up with some ideas, discuss them, and then rule out the bad ideas.

It is fairly odvious to me that the Bot is open-ended.

Most likely V1 will be a very simple product, Tesla is likely to just get something working then refine it.

 

[MC3OZ]

To further explain attachment points, a close parallel is lego.

Attached arms would be standard and a number of different tools could go at the end of the arm.

The bot could be configured for its role on the production line.

The ideal would be to set the bot up to do a great job of its role on the line. It then just does the job 24x7x365 whenever the line is running.

Elon may bbe right about needing a UBI, but if the economy expands on the back of cheap energy and labour that might create a sufficent number jobs. Through out history innovation and higher productivity has created new types of jobs.

 

[MC3OZ]

Bot info from AI day:-

• Weight 125 LBS
• Human level hands
• 40 Electromechanical Actuators
• 2-Axis feet for balancing with force feedback sensing

Seems like Tesla already has a fairly specific bot design in mind with a focus on human hands. So we can rule out attachment points for now.

Perhaps attachment points are like Lidar, and human hands are like pure vision. Tesla isn't a start up in need of quick results.

However, temporary shortcuts are available if they ever need them.

 

[MC3OZ]

For the battery pack, ideally they don't want a cooling system in the bot, which suggests LFP.

125 LBS = 56.7 kg

Rough ballpark maths of LFP at 180 Wh/kg

• 3 kWh - 16.7 kg
• 5 kWh - 27.8 kg

The actual weight of the battery pack will be higher and some part of the pack might be a buffer.

3-5 kWh seems like the right range with the lower end being more likely.

But if we take 5 kWh of LFP at $100 kWh - we can see that the bot battery only costs around $500.

Actuators range in price from $10 to $2,000+., Tesla would be doing a bulk buy, or building these in house, so I would guess around $50 each $2,000 in total

If anyone has a better guess, please post it.

So a $10,000-$15,000 hardware cost, still seems possible.,

 

[heltok]

Do you guys really believe it will cost ~the same to make the robot as to make a Model 3? Elon would never accept this, he is probably aiming at same cost/weight as the Model 3 plus some non weight dependent components such as HW4, cameras etc. I’m gonna guess materials+manufacturing at $3k/unit once they reach millions/year.

Spot weights 30kg and his competition is selling in China for $2700 today:

This Chinese robotics startup is selling a knock-off of Boston Dynamics’ Spot robot that costs $75,000 for only $2,500 - Tech Startups

Shanzhai is a Chinese term that literally means "mountain fortress" or "mountain village." But the contemporary use of the term usually encompasses counterfeit, imitation, or pirated products, which can involve the imitation and trademark infringement of brands and companies. You may argue that...

techstartups.com

 

[MC3OZ]

I’m gonna guess materials+manufacturing at $3k/unit once they reach millions/year.

Not a bad guess, I'm thinking around $10k/unit, but my guess might be on the high side.

I also think the bot can be designed for easy assembly/disassembly, cast frame (skeleton) sections with locating pins, bolted together.
The outer skin sections are vinyl covered foam, similar to car seats.
Most of the rest is circuit board and wires, + a screen - ball park for this probably similar to the cost of a desktop computer.

So while I think your guess is aggressive, I would not rule it out, especially with Tesla working to progressively reduce costs.

I've previously guessed that up to 1 Million bots per year is easy to do.

We need to distinguish between some of the software and hardware engineering being difficult and iterative, and the cost of building the hardware which is often a fairly fixed cost, even when upgraded versions are made..
The best way to effectively recoup R&D costs is, high volume manufacture.

And the best way to do training is, with a large fleet.

 

[EVCollies]

WOW, great ideas @MC3OZ !

Would you need mounting bolts on the bot or how about a backpack?

Instead of a specialized bot for a function, I would imagine a specialized tool for the generic bot to accomplish that function.

Elon is brilliant to start with a biped robot as our world is currently optimized for such a bot. These will be much more utilized than a dog or a Roomba-like bot that can work in a specialized environment such as sorting warehouse.

 

[MC3OZ]

WOW, great ideas @MC3OZ !

Would you need mounting bolts on the bot or how about a backpack?

Instead of a specialized bot for a function, I would imagine a specialized tool for the generic bot to accomplish that function.

Elon is brilliant to start with a biped robot as our world is currently optimized for such a bot. These will be much more utilized than a dog or a Roomba-like bot that can work in a specialized environment such as sorting warehouse.

I guess my main point isn't Tesla is strictly limited to a complete replica humanoid form, they can enhance the form when needed.

if we imagine the bot as a cast frame with wires running inside, it is possible to create slots where additional accessories can be mounted and can plug into power sources. Once they have an extendable platform, there are very few limits to where it can go,.

I see accessories as docking and locking into the structure, using power supplied by the bot to move the locking mechanism, reverse for undock.
When unplugged the accessory has no power, i.e. no batteries in the accessory.

For gear on the back, it could use a backpack.

But when I revisited the AI day presentation, it looks like Tesla are going to make the best possible humanoid form, and determine what they can train it to do, So all of my speculation on accessories is off track for now.

I think the economics are also likely to be great.

My estimate (at scale) $10K hardware, $10k software and training operational life with servicing 20 years, servicing average $200 per year.
All up cost for 20 years:- $24,000 - able to work 24 x 363 (can charge at work site - 2 days off for servicing)
Ignoring electricity costs - $1,200 per year, $3.31 per day, $0.14 per hour.

 

[MC3OZ]

Another way of docking an accessory tool is, if the bot hand fits inside the tool adaptor, like a glove, that process could establish a power supply connection

Again that is a standard where multiple tools (adaptors) can be built specifically for the bot.

This differs from the bot simply holding to the tool in its hand in the following ways:-

1. It is possible for the bot to provide power to the tool.
2. The bots grip on the tool is always precise and firm, within very fine tolerances.
3. It is impossible for the bot to drop the tool, pick it up incorrectly, or orientate the tool in the wrong direction.

Hand in glove is probably superior to attachment points, it achieves most of the same aims in a more compact design with less complexity.

 

[mblakele]

Covid has reset relations between people and robots
Covid has reset relations between people and robots
from The Economist

Excerpts:

In the field of manufacturing, car plants lead automation. But, as Henrik Christensen, director of the Contextual Robotics Institute at the University of California, San Diego, told the meeting, even the most advanced of them, those in South Korea, average only around one robot per ten workers. So-called “lights-out” car manufacturing, with no human beings on the factory floor, remains a distant dream.​

​

[…]​

​

Robots that work with people in such ways do, though, require special training. And there is a long way to go to improve that says Julie Shah, who leads the Interactive Robotics Group at the Massachusetts Institute of Technology. Most robots perform narrowly defined tasks, with mobile ones using their sensors to avoid bumping into people. “Robots need to see us as more than just an obstacle to manoeuvre around,” adds Dr Shah. “They need to work with us and anticipate what we need.”​

​

Studying what happens in factories shows that the most successful applications employ robots programmed by an engineer who is working side-by-side with someone (a so-called “domain expert”) fully versed in the tasks at hand. To make that easier, she and her colleagues are developing ai systems which can school a robot using natural-language commands.​

 

[betstarship]

Covid has reset relations between people and robots
Covid has reset relations between people and robots
from The Economist

Excerpts:

In the field of manufacturing, car plants lead automation. But, as Henrik Christensen, director of the Contextual Robotics Institute at the University of California, San Diego, told the meeting, even the most advanced of them, those in South Korea, average only around one robot per ten workers. So-called “lights-out” car manufacturing, with no human beings on the factory floor, remains a distant dream.​

​

[…]​

​

Robots that work with people in such ways do, though, require special training. And there is a long way to go to improve that says Julie Shah, who leads the Interactive Robotics Group at the Massachusetts Institute of Technology. Most robots perform narrowly defined tasks, with mobile ones using their sensors to avoid bumping into people. “Robots need to see us as more than just an obstacle to manoeuvre around,” adds Dr Shah. “They need to work with us and anticipate what we need.”​

​

Studying what happens in factories shows that the most successful applications employ robots programmed by an engineer who is working side-by-side with someone (a so-called “domain expert”) fully versed in the tasks at hand. To make that easier, she and her colleagues are developing ai systems which can school a robot using natural-language commands.​

Reminds me of this article I read awhile ago:

'Golden era of robotics adoption' kicks off in 2022, strategist says

Next year will be a "key inflection point" for growth in robotics and artificial intelligence, says Global X's Jay Jacobs.

www.cnbc.com

 

[Buckminster]

I bet it'll be 2.557 kilowatt hours (rechargeable in 20 min @ 3C)

That's one human liver's worth of glycogen energy (about 2,000 calories) plus 10% extra capacity for charge buffering.

Optimus' should beat humans handily on energy efficiency though (similar to ICE vs EV), and the AI CPU uses about as much power as a human brain (100w).

Once trained properly, I expect Optium' will be about 3x as productive as a human laborer, and able to work 21.5 hrs per day. With vacations and time off, that one 'bot replaces 6 workers. With its productivity, it creates the value of 18.

Cheers!

 

[Buckminster]

It's incredibly ambitious and optimistic. If Tesla can get an Optimus proto by the end of 2022 and any at all Optimus production by the end of 2023 I will literally be sitting with my jaw open in shock and awe.

I'm not saying it's impossible because Tesla constantly amazes and surprises me, I'm just saying the timelines seem a bit too optimistic for me to think they can truly achieve this that soon. That would mean we'd see Optimus in production before the $25K Tesla EV. To me that seems very unlikely?

Like others have mentioned I would compare this venture into robotics to SpaceX and reusable rockets. If NASA couldn't do it what makes Elon and Spacex think they can accomplish this seemingly insurmountable task?

The answer is Tesla's (and Spacex's) culture of innovation. In 2014 I had a conversation with my cousin's husband who at the time was working at SpaceX. I was blown away when he told me the version of the rocket may not change, but the components were constantly changing.

New components were integrated on an almost weekly basis. This was a large part of my investment thesis and reason I began to open a position in Tesla knowing they would have an absolute advantage when compared to the molasses slow innovation rate within the auto industry.

Without getting into specifics I had the opportunity to work with the Boston dynamics dog and run it through its paces as an end-user. Pros and cons to it compared to other robotics we were using, but any "learning" was not done in real-time. Walking was very good especially with obstacles and uneven terrain, but not perfect and we were able to find areas it couldn't traverse like certain hilly terrain. Opening doors was ok and certain knobs and configurations confused the robot.

Tesla's ability to innovate extremely fast, and Teslas AI will rapidly set the Tesla Optimus bot apart from any offering from Boston Dynamics. Prototypes are easy; production is hard. Once Tesla bot is ready and available, it will be an absolute game changer in the industry I am in. A proverbial paradigm shift.

Elon has mentioned recently and on several occasions that people will soon see Tesla as a robotics and AI company. I have no reason to doubt him, and yet my projected valuation of Tesla doesn't fully factor FSD, robotics, or AI. The reason: the numbers are just nutty.

Actually... The mechanical parts of Optimus I don't see Tesla having any major issues with. Lots of dedicated, enthusiastic ME, good large scale manufacturing experience, etc. Again, don't see a problem.
The real problem with Optimus is the smarts. We're autonomous two-legged, two-armed entities that can roam the landscape and do multitudinous things, but we've also got ~3 billion years of evolution behind us. Things that didn't work didn't survive. Things that did work did. Hit or miss, lots of failures, but lots of time to get it all working.
So, how the heck does Tesla get something (ha!) simplistic like an autonomous car driving around? Answer is: No-kidding big neural networks (that's how we work..) and training, training, training. A lot of our training is built in; lots more is learned (nature vs. nuture), but, either way, it's in the wetware in our skulls.
The Thing That's Coming: Dojo. Dojo is a supercomputer for training neural-network based AIs. There's nothing like it on Earth at the moment; Boston Dynamics, despite their prowess, has nothing like it, and Tesla & Co. is bringing Dojo on line this year. They Have Plans for Dojo, the beginning of which is to flesh out FSD. But once they get past that: Well, one has invented a machine that can train other machines, and it's generic. And this is where Optimus will come into its own. Simple tasks by a human's lights, but capable of running, say, a paint shop, with oddball things happening.
First get FSD going; then switch gears for Optimus; and then make the Dojo platform available for others who want that kind of training. It's going to be a new world.

I have a little background in robotics, taken quite a few courses in mechatronics and built a few robots at the university but not in the industry.

Boston Dynamics are very good at pushing software 1.0 to the limit for 2 and 4 legged robots, but they use close to zero deep learning, at least they didn’t a few years ago. The rely on Lidar to figure out how the world looks like and then use classical control theory to move it around in the world. Sometimes they use humans to do acrobatics and then tell the computer to calculate the control signals to repeat the same sequence of reference poses. Nothing super hard that Tesla couldn’t do but it sure looks fancy.

The hardware is nothing special either. A few electrical motors that Tesla probably are even better at than BD. A battery that Tesla obv know well. Some sensors where Tesla are expert at cameras while BD uses Lidar. Some compute where Tesla have very optimized propritary hardware already and can probably just slow down to save energy if needed. A skeleton that needs to be propotyped that Tesla can do very well. So I don’t see where Tesla would be lacking.

Tesla have been doing robot prototypes for a long time. The snake robot was 6 years ago:

And they have probably recruited several engineers who have built project such as this 7 years ago:

And after AI day, most of the people in the people in the competition will be having wet dreams about joining Tesla…


So yeah, I think it is likley that Tesla in fact will have a prototype around the end of this year.


Not sure how many workers the robot will be replacing in Tesla’s factory. If a optimus can do a task, then a kuka robot can probably also do it, but much more precise, fast and efficient.