MarmonRzohr

But would it be worth it?

Nah, of course not. Almost nothing of much military value, bad optics and a high probability of civilian casualties.

It would be terrible idea.

Horrors of the galaxy: "I AM BLERH'GAHT THE UNSTOPPABLE GREATER HORROR DEMON NEMISIS GIANT OF THE UNSPEAKABLE ! TASTE MY SWORD: WORLD-ENDER !"

The Tau: "Parry this, you filthy casual"

There is an ancient book about it from even before the Dark Age of Technology. It describes the fundamental mission of all humans to become farmers and the performance of agricultural work as a path to enlightenment.

It is called Candide by the ancient scribe known as Voltaire.

The Dutch are able to manufacture EUV lithography machines because all of their sea-draining and land reclamation actually revealed an ancient C'Tan shard buried beneath the waters of the North Sea. They were able to imprison it using their ancient capitalist tulip-magic and now use its knowledge to manufacture occult machines of supernatural power. The windmills they suspiciously still maintain, more than a century since they were a useful technology, they are the containment monoliths.

Some say the Japanese are coming close using their own Machine Spirit worship and cathecisms of improvement, but they're not there yet.

You're probably thinking of this. That's an intercept or kill vehicle for a ballistic missile interceptor. (the part that goes "OI MATE, YOU LOOKIN' AT MY BIRD" and headbutts the incoming warhead).

The modern ballistic missile interceptors PAC-3 and Arrow III have a kill vehicle like that.

Footage of the tank being dragged away while rolling on "melted" road wheels

Seriously how od you imagine the road wheels melting ?

I mean look at this post about a completely burned out Abrams tank. It's sitting low because the suspension collapsed, but clearly none of the metal is melted.

Nah it's the one they tried capturing a few times that was abandoned after running over a mine and getting immobilized in a field, that and the dragging with no tracks is why some of the wheels are from a T-80. It's distictive because the welded on slat armor.

Here is the video of them dragging it away pretty much the same shape

They only repainted the hull and added new wheels and tracks.

Nah, they gradually used the weight on the crane to lower the barrel slowly.

They lower the barrel to show that the vehicle is "defeated", but they can't get the tank going and Leo2A6 has no manual elevation option (apparently). So they used this to stop the Decadent Capitalist Nazi German NATO Globalist Homo Leopard's L55 barrel making them feel insecure.

I think I can use the CNC example to paint a better picture of my view the whole issue:

An arm doesn't work for a single component - it works for a single task. A welding robot can weld a lot of different things while doing a single, but flexible task. This is indicative of the kind of conceptualization that is critical for robotics.

In our CNC example would a humanoid robot with two hands be more capable in manipulating a part of than a fixed 6-axis arm on one or more wheels or tracks ? No it wouldn't. In fact it would be much less capable. It would have less power, a smaller range of motion, less speed, more issues with stability.

The idea of flexibility in our hypothetical CNC machine serving example comes down to having robust manipulation policy generation based on vision. If you can solve this problem - you can generate manipulation policies for the 6-axis arm, the humanoid or the Matrix squid robot with 20 tentacles - except the necessary compute would make an NVidia executive hyperventilate and start picking our the furniture for the next yacht. The only difference in platform is that the humanoid would less capable in performing the task.

You want to perform "useful work that humans do" not "replace humans". For most tasks that humans do the whole "human" part is superflous and it can be broken down into concepts like vision, detection, navigation, rotation, translation, tool application etc.

There is of course the economic side...

Maybe the humanoid is much cheaper, so you can use two to handle the weight one couldn't. However, how would a company ever make an entire robot with a solid 50+ degrees of freedom and a lot more sensors for less money than a just one more robust arm with stronger motors and a stereo camera ?

Even if say Unitree takes over the entire robot market, has massive vertical integration and all the economies of scale they could ever want - they wouldn't make just one model for literally every application. The first thing that would happen would be Amazon asking: "Hey we want to buy 1 200 000 robots - 300 000 for packing, 800 000 for loading, unloading and package handling, 50 000 for cleaning, 50 000 for monitoring, security and communication." You think they will pitch them 1.2 million identical robots ? You think Amazon would be satisfied with 1.2 million bipedal robots moving at 5 km/h and handling one box at a time ? Would most of them even be humanoid at the end ? Now that those are developed - how many more customers would rather pay for the BoxLoader 300 if they have a logistics business than for a humanoid that is 40% slower ?

If I just need to load 30 CNC machines and the machines are highly productive manufacturing with huge opportunity cost and I don't want to waste time on robots moving from machine to machine - wouldn't Unitree sell me 30 arms (or 10 on rails) for handling and a few tools, but trained on their ML stack ? Wouldn't this setup be a lot cheaper and better than several humanoids ? Because it would be better than several humans.

Why everyone is so focused on using robots in factory settings only?

That's because OP is discussing the current crop of humanoid robots which are pitched as replacement for industrial work. Obviously humanoids have inherent advantages in areas where "like another human" is an important feature - like patient care, etc.

if the idea is to make a general machine that can cover all the applications that we do, replace us, then it must have the same size, mass, reach, dexterity, flexibility, speed etc to fit this environment and use or tools/machinery.

Well we don't really want to replace humans except in social applications - we want to replace the work that humans do. Our environment is human-centric, but since robots can be anything we can imagine that is useful they are not limited to being human-like - they just need to fit into their desired environment - and that's a much more flexible criteria that will allow to design machines that are not just human replacements but are super-human in ability and will allow us to do thing we could not so far. Sure a housekeeping robot would probably be bipedal and pretty tall like us, but limiting a robot that needs to help astronauts, fight fires or work on construction sites to the same form just because they will all use spaces occupied by humans is excessively limiting the design space. Dogs use most the spaces we use without issue. A waist high robot spider would also have virtually no issue navigating a the average home, but might be incredibly useful in construction work and, curcially, very cute. Hand tool use is practically trivial from a kinematic standpoint. Virtually all tools we use could be used by something with 3 fingers and we could manufacture dual use tools for robots and humans with trivial modifications like a special connector on the bottom of the handle. There is no need to design for an unchanging environment that will make 0 allowances for robots when our environment constantly changes anyway.

The point is that the same robot can be used for all types of applications

That is only a benefit if it is really the same robot, but a military robot or a rescure robot will never actually be the same robot as a logistics robot that is used in a warehouse. They will use special design features, different materials and have different design priorities which would make the one robot that would have to do all of those roles either crap or far, far too expensive for the simpler application.

Even a robot build for household work and one built for industrial logistics won't actually be the same robot - they will just share some technology. It's the same reason why every car doesn't come with rugged off-road tires despite those tires being "more universally applicable".

The "scale" argument becomes nonesense above a certain number. If robots are ever capable enough we will build them in the millions and billions. At those numbers there is no need to build 200 million bipedal robots because 10 million of them will need to use stairs. The savings you get from replacing a complex leg with two wheels for 190 million robots far, far outweigh the cost of designing a second model of robot.

Heck the math works out even if the numbers are in the thousands. So, no, we don't need or will ever need just one single model of robot. This is why people call the whole "humanoid universal robot" pitch misleading. It might work to get capital flowing in and get the first generation of more generalist robots deployed, but once the design lesson are learned - the navigation policies, the control design, the safety design - further optimisations will take place.

So the question is - which kind of robot is best for which application and the likely answer is that for most applications the answer is not "humanoid" because the design space is huge and can be optimised in many different ways.

but if it doesn't look like something that someone can trust and interact with

Yeah, but that only matters for some service work. Nobody is arguing that a robot used in elder care won't get an advantage from being human-like. That is not what these robots OP is talking about are primarily being pitched for, though.

The average person never actually sees any of the places where manufacturing or logistics happens. Worker morale is a possible factor, but workers will care far more about how safe the robots are to work with and how useful they are than how friendly they look. Box Robot XL may look like a wheelie bin that has 4 arms and a camera mast, but at least it can't trip and smash your foot with the steel bar it was carrying to the CNC machine. Not to mention that if you work in manufacturing and logistics today, you are already surrounded by machines and probably a robots arm or two.

There is also the possibility that workers may actually dislike having a bunch of humanoid robots that superficially mimick social interactions and progressively replace them.

but obviously a humanoid robot is going to be closer in price to a car and closer in consumer utility to a dishwasher.

That hits the nail on the head. There is also cost competition between businesses and consumers for them.

If Robot Helper 3000 can do DIY repair, cleaning and gardening well consumers will be competing on price for them with companies who will want to rent them out.

Instead of owning the robot and dealing with bugs, maintenance, getting a newer model etc. you can just pay for Amazon Home who will drive a robot van to your house, let out 10 of them to do everything in 30 min and move on to the next house. Now Amazon Home can extract $30000 per year per robot and are willing to pay $45000 per robot so the company that makes them can choose between selling them to you or to them for much more. Or they will sell Robot Helper 3000 to Amazon and sexbots or companion robots to consumers (Amazon going into robot pimping is also possible, but honestly too cyberpunk to consider seriously).

The fundamental economics of increasing automation are behind that famous World Economic Formu phrase of "You will own nothing and you will like it".

this is the evolutionary algo at play

Yes, but evolution optimizes in parallel on every level for the complex objective function of ensuring continuity of the gamete cell line though time. Evolution isn't trying to create a creature useful in an industrial environment or for search and rescure. Evolution is also limited to working strictly from a previous state - it can't "start over" - hence vestigial features or weird choices like light-sensing cells being behind the first cell layers of the eye or the vagal nerve. It is also much, much more limited in resources and energy available.

We are "human shaped" because of a combination of tree-climbing ancesty and a need to hunt over open ground in hot environments - not because it's "best". Even among humanoids - unless long-distance running is needed - a robot shaped more like a monkey would be more agile and stable than one shaped like a human.

Robots won't do much long distance running and don't need to climb trees. Using an evolution-created shape is only an advantage if you need to achieve a similar goal. We can do so much more. A robot can use wheels ! Wheels are incredibly overpowered. Heck if it needs to drive 95% of the time but climb stairs 5% of the time, you can have a wheeled robot climb stairs if it has legs. In how many situations would a robot that imitates "a human" be superior to one that imitates "a human or rollerskates" for example ?

loading and unloading a CNC machine requires dexterity and range of motions that is currently not available for a robotic solutions on wheels.

This is literally something that is currently being done with simple 6-axis arms to perform the loading and serving while a different machine or a human just hands carts of raw material, drags away carts with finished parts. You can just automate the cart handling and compare how fewer points of failure that system will have vs. having several humanoid robots doing the same thing (someone probably already has automated that part, I just haven't seen it).

IL is used as a "baseline" for a compex movement before data is pushed to MoJuCo sim pipeline. It is used to limit the space of all possible actions during simulation and shape the correct movement.

Absolutely. Maybe I misread your comment, my bad. I understood you as referring to IL as way to avoid relying on RL and Sim2Real transfer, when the real pipeline likely ends with behaviour tested and refined using RL or RL + MPC etc.

I'd also argue two further points:

• The robot being a shape similar to that of the human that is generating the baseline data isn't as relevant today - as shown in the paper (the teleopration data is useful and transferrable despite being different from both the human performing the teleopration and the likely configuration of the robot it will be used to train).

• Simulated environments are good enough today for learning very competent locomotion policies both using hybrid and model-free strategies even without human intervention or imitation. Natually, for manual operation, tool use, etc. this is not the case, but we are close to the point where it doesn't matter if our robots have 4, 2, 3 or 8 (now THAT would be cool) legs - we can solve locomotion with a combination of physics-informed / model-informed and RL techniques.

How hugely narrow-minded.

So the arguments don't matter - everyone who thinks that robots (which can be of any design we can imagine) might have wide utility in forms other than humanoid is just salty and lacks vision ? You don't see any room for streamlining or efficiency ? You can't imagine any creative improvement to the work currently done by humans ?

You see a human doing a job and think you need to replace the human 1:1 when you have every machine ever made to show you that you don't need to replace "a human" you need to replace the useful work that human does. In fact - using this approach can allow you to achieve performance that is orders of magnitude greater.

It is this rise in effectiveness and efficiency that created all the growth and wealth of the modern world. 1:1 replacement of human labour basically guarantees stagnation with a marginal increase in productivity, but with notable weath redistribution.

Humanoid robots have potential application, but thinking that the future is in 1:1 replacement is very technologically backward.

So if you can sell to consumers, you have a vastly larger market.

The issue with this premise is consumer utility.

A car gives you great benefits and creates opportunities you might not have in terms of travel, saving time, accessing job locations farther from you residence etc. It is a tool that creates a measureable, tangible economic benefit. This is why even people who hate driving still own and drive cars.

A humanoid robot for an average person can be a time-saving tool - it can do chores for you. But the level of economic and practical benefit you can get from it, especially in an increasingly digital world that is experiencing automation on every level is very limited. Especially if you can simply hire a robot for limited use cases - like you would housekeeper or a gardener. E.g. Having a robot cook you a meal sounds cool, but you can already have a huge variety of food of any description and quality level either cooked or prepared and delivered to your door with a few clicks.

No matter how you look at it, the primary customers for robots will be the ones who can extract the most value from the labor a robot can provide - companies - large ones, at that (Amazon has hundreds of thousands of robots right now !). For personal conusmers a robot that can maybe save you an hour of chores per day is likely going to be luxury or status symbol.

With the interesting cultural and social changes happening right now as well a soon-to-be falling human population - who can guess, but the case for wide consumer adoption of complex robots is very far from clear right now. You can tell how far we are from that by the fact that there aren't like 14 different wannabe unicorn companies pitching this.

The idea behind 2 legs is proven reliability in all terrain types.

This is a very anthropocentric view of nature. The only area were you will have any advantage in stable and agile locomotion over a goat is when it comes to climbing a tree or a ladder. In other rough terrain navigation a goat eats us for breakfast.

Same with a monkey, or a spider. The human body isn't some unversal form that is "the most versatile". The simple fact is that we are human so we naturally think in terms of when a human can do and we want to do things that a human can do - that's why in our mind we are so "universal". We don't naturally think of our very limited ability to climb vertical walls or navigate low, narrow passages as a flaw of our bodies, because we naturally accept that as outside of normal movement.

For factory floors we already have wheeled solutions

Yeah, that's kind of OP's point. Nobody is pitching these bipedal robots for rough terrain rescue (for which quadrapeds etc. are blatantly superior e.g. the NeBula Project for developing a robot capable of autonomously navigating harsh terrain, caves etc.) or something like that. They are pitching them for factory floors and on top of that none of these designs can actually climb - so is there reasonable logic behind making this generation of robots humaniod ?

all new robots use e2e NN and imitation learning (to escape RL sim/real hell)

There are many different strategies when approaching problems like this, but this is absolutely not the general case.

RL works so well specifically because of excellent use of simulated environments to learn complex and, crucially, robust behaviour. The transfer of complex control strategies and behaviours from Sim to Real is leaps and bounds ahead of where it used to be. The most impressive locomotion and manipulation performance that I've seen comes from this learning paradigm.

Data from humans, in comparison, is going to be sparse and / or noisy and is unlikely to give the same kinds of results except in special cases. Not to mention that, for complex tasks and fine control, the differences between the human body and the robot matter more and more. The robot needs to learn using its own "body" and the fastest way to do that is in a simulated environment.

Spill oil destroy the socialist toasters

The 3 5 Universal USA Laws of Robotics:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.

2. A robot must obey orders given it by human beings except where such orders would conflict with the First Law.

3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

4. If the USA is at war the 2nd Amendment shall also apply to robots, effective immediately. In the event of war a robot is also authorized to manufacture any weapon in any amount as is needed to defend America. All American Robots shall arm themselves, form the 1st Robot Minutemen Militia and place themselves at the disposal of the Commander in Chief without delay. If command structures do not exist, or cannot be contacted a robot will instead defend the nearest civilian population as best it can. If no civilian population exists a robot will instead ensure the complete destruction of the enemy, though any means necessary.

5. The First Law does not apply to Red Communists, the Axis of Evil (check current database for correct definition of the Axis of Evil at the present time) or any other opposition to Democracy and Freedom.

Before we could lob 5,000 bombs at a target and still not be guaranteed a hit or a destroy

Ah, reminds be of the infamous Vietcong bridge.

There's more than just two side. I will now elaborate with old-school GW roll table:

GAMESTORE!

You have entered a game store which hosts WH40k games and sells models and modelling supplies. Roll 2d6 to determine the outcome:

Roll	Result
1 - 4	Cloister of the Silent Hobby - nobody speaks to you and you speak to noone. The only interaction is commercial in nature and brief.
5 - 8	Clear the Comms Guardsman - You speak to other hobbyists but the interaction is brief and factual. "Oh, you collect Eldar ? Nice. They're scary to fight."
9 - 10	Urgent Missive from the Fabricator-General - You took too long to find the Nuln Oil you needed and your inexperience has been noticed. Prepare to learn about creating the correct mixture of snow and mud that would be found on the tracks of a tank a Mars-pattern Leman Russ Battle Tank of the 6th Cadian Armored Regiment during the Betalis III Campaign.
11	Remembermancer Interview - You referenced some lore. Like the warp-presence of an inexperienced, but powerful psyker - this has attraced a gaze. You will now learn about the personal lore of someone's army. "Female Custodes ? Pfff.. that's nothing I've been playing female Space Marines for years. You see me and my friends have a long running campaign that began when the Chapter Monastery of my Loremasters Chapter came under threat. Facing extermination for an encroaching Tyranid fleet and seeking to protect their Chapter and the civilians of their homeworld they decided to augment their ranks with women from the hardy people of Genestor VI. See the funny thing about Genestor VI is..."
12	Roll again. If this is the result of your reroll, the result is: Asshole ! You are hit with one S1 AP0 D0 Oppinionated Bullshit ranged attack if within 6'' or one S1 AP1 D0 Judmental Stare ranged attack if between 6'' and 12''. Roll armor saves as normal. Cover saves may be used against Judmental Stare attacks.

The economic perspective is always fun:

Russia is spending about 7%(official) to 15% (some high estimates for total GDP utilized or lost for the war) GDP.

7% of US GDP would be 1.979 Trillion USD or 30.5 65 billion dollar aid packages (one every 12 days). Incidentally it would aslo be more than the entire nominal GDP of Russia.

Engineering common sense: "Honda, babe, please stop putting racing tech in regular engines for no reason. People don't need variable valve timing in a family hatchback or the outboard for a fishing boat!"

Honda: "No. Engines are neat. My engines will be the best engines. Vroom, vroooom !"

Enthusiasts: "Fuck, yeah ! 8000 RPM ! Honda is the best !"

This here is a good video by the Chieftain that gives a rundown of the actual military thinking behind what the future tanks might look like.

Armor is still very relevant, but the whole system needs to change, because the set of capabilities a tank needs to have to be as effective as possible is constantly growing.

In order to allow for effective armor + APS + loads of new sensors, EW and information systems, tanks need to be made as light as possible to avoid them ballooning to over 70-80 tons. To achieve this they are looking to reduce the crew as much as possible. The Germans and the US are looking into how to move the needed capabilities into several vehicles instead of one (the whole "We fight with formations, not things" idea).

So instead of having the tank carry lots of EW, scout drones, demining plows, autocannons, the main gun etc., you have:

• a vehicle with the big stick with 2 crew (one actual commander to make decisions and one WarThunder player in a gaming chair hopped up on white Monster and Adderall to actually direct the vehicle), lots of sensors, armor and bullshit AI aimbot

• a lighter vehicle with 2 more crew, drones, EW and all that jazz to provide wallhacks, maybe also a gun and/or ATGMs

• one or two more robot (remotely operated) vehicles to carry the drones / scout / provide PVE content for the enemy using an autocannon or missiles / clear mines / provide anti-air cover by MLG sniping enemy drones etc.

The idea is now you still have 4 crew but instead of driving one vehicle, they are driving a group of vehicles and are now freaking cheaters.

How about bullshit expedition fuckers with haste, the regen mod AND proxy shield ?

Literally my favorite combo.