Humanoid Robots for Moon and Mars

[sanman]

Humanoid robots have a data problem.  There isn't enough to feed re-enforcement learning. Many of these companies are using Nvidia Cosmos to generate that data through hundreds of millions of physics based simulations.  I can't help but wonder how that works for Clone.

Modelling and simulating a robot with tens of actuators is one thing.  Modelling hundreds of squishy and wobbly hydraulic 'muscles' sounds like quite another.  Even if Cosmos handles it like a champ, there's presumably a higher computational overhead, making training slower and more expensive.

ASICs can handle it. There's a reason Broadcom stock started shooting up in Fall of 2022 alongside Nvidia.
Broadcom is the biggest implementor of custom ASIC technology, and have the most industrial expertise in it.
Whatever you can do with software, you can hard-code it in hardware thru ASICs.

[BN]

https://en.wikipedia.org/wiki/Experience_curve_effects

More things made = better things.

Well thanks for coming up with you own answer for the question I posed to member sanman, but what you reference applies to the COST of something being produced, not the capabilities of the product. So not relevant.

You're being purposely obtuse. 

So no, building more of something does not mean that something will gain new and better capabilities. In fact you don't even have to build something in order for the design to iterate and get better - this happens all the time in engineering departments.

Also, you are assuming that the mechanical parts of a humanoid robot are able to iterate quickly, despite decades of robotic design and manufacturing pointing to the opposite.                     

in this case, scale = more real world feedback
real world feedback is the limiting factor beyond modeling and simulations.

the mechanical parts of modern robotics are being improved very quickly. this is apparent to anyone paying attention.

You can have the best end effectors, but unless you have a control system that can use them to accomplish a task, they will be useless. And that is what we have NOT seen, in any humanoid robot, is the ability to do work that comes close to what a human can do.

[BN]

simple and proven

[daedalus1]

Lol.
There will soon be no reason for humans on Mars. The few that are there will be looked after  by humanoid robots. Much cheaper and only a one way trip required.

[Cheapchips]

Humanoid robots have a data problem.  There isn't enough to feed re-enforcement learning. Many of these companies are using Nvidia Cosmos to generate that data through hundreds of millions of physics based simulations.  I can't help but wonder how that works for Clone.

Modelling and simulating a robot with tens of actuators is one thing.  Modelling hundreds of squishy and wobbly hydraulic 'muscles' sounds like quite another.  Even if Cosmos handles it like a champ, there's presumably a higher computational overhead, making training slower and more expensive.

ASICs can handle it. There's a reason Broadcom stock started shooting up in Fall of 2022 alongside Nvidia.
Broadcom is the biggest implementor of custom ASIC technology, and have the most industrial expertise in it.
Whatever you can do with software, you can hard-code it in hardware thru ASICs.

Tensor processing units, like the ones Broadcom make for Google, are not 'hard-coded' for AI. TPU's specialise in high volume/low precision compute, which is what you need for training or executing AI workloads.

Running a walking or grasping or vision policy on a robot is always going to need that specialised but unfixed compute. With current approaches at least.

[catdlr]

simple and proven

That picture assumes that China arrived first on Mars and the Robot is Chinese-made. 

[edzieba]

simple and proven

And in direct competition with a single driven front wheel and a steering linkage, to go with the two existing back wheels.

[Twark_Main]

Humanoid robots have a data problem.  There isn't enough to feed re-enforcement learning. Many of these companies are using Nvidia Cosmos to generate that data through hundreds of millions of physics based simulations.  I can't help but wonder how that works for Clone.

Modelling and simulating a robot with tens of actuators is one thing.  Modelling hundreds of squishy and wobbly hydraulic 'muscles' sounds like quite another.  Even if Cosmos handles it like a champ, there's presumably a higher computational overhead, making training slower and more expensive.

ASICs can handle it. There's a reason Broadcom stock started shooting up in Fall of 2022 alongside Nvidia.
Broadcom is the biggest implementor of custom ASIC technology, and have the most industrial expertise in it.
Whatever you can do with software, you can hard-code it in hardware thru ASICs.

Tensor processing units, like the ones Broadcom make for Google, are not 'hard-coded' for AI.

Nobody mentioned TPUs. Sanman mentioned ASICs (Application Specific Integrated Circuits), which is a general term that does include chips "hard-coding" for AI.

I'm guessing the mention of Broadcom (as an illustrative example of the recent ascendancy of ASICs) caused some confusion.

[Cheapchips]

Humanoid robots have a data problem.  There isn't enough to feed re-enforcement learning. Many of these companies are using Nvidia Cosmos to generate that data through hundreds of millions of physics based simulations.  I can't help but wonder how that works for Clone.

Modelling and simulating a robot with tens of actuators is one thing.  Modelling hundreds of squishy and wobbly hydraulic 'muscles' sounds like quite another.  Even if Cosmos handles it like a champ, there's presumably a higher computational overhead, making training slower and more expensive.

ASICs can handle it. There's a reason Broadcom stock started shooting up in Fall of 2022 alongside Nvidia.
Broadcom is the biggest implementor of custom ASIC technology, and have the most industrial expertise in it.
Whatever you can do with software, you can hard-code it in hardware thru ASICs.

Tensor processing units, like the ones Broadcom make for Google, are not 'hard-coded' for AI.

Nobody mentioned TPUs. Sanman mentioned ASICs (Application Specific Integrated Circuits), which is a general term that does include chips "hard-coding" for AI.

I'm guessing the mention of Broadcom (as an illustrative example of the recent ascendancy of ASICs) caused some confusion.

The chips that Broadcom build for Google are classed as ASICs. They're TPUs though.  They're not a fixed silicon implementation of a piece of software, which was where I felt Sanman was pointing. They're only application specific in that TPUs are optimised for they type of calculations required for running AI in general.

[MickQ]

Lol.
There will soon be no reason for humans on Mars. The few that are there will be looked after  by humanoid robots. Much cheaper and only a one way trip required.

There will always be a reason for humans on Mars because we are Human.

[BN]

simple and proven

And in direct competition with a single driven front wheel and a steering linkage, to go with the two existing back wheels.

that is called a tuk-tuk. not ideal for off-road and would require motor, batteries, bms, ecu, steering, etc. and it could break. it's complex enough that you might as well bring a proper car.

the rickshaw is just a seat on wheels with two handles. light, effective and reliable. even so, it might be a dumb idea but kind of interesting to think about. if an astronaut was injured, a robot could chuck em in this seat and bring them to base. better than a stretcher or a litter which would require two robots. 

simple and proven

That picture assumes that China arrived first on Mars and the Robot is Chinese-made.

we will prevent this with mars tariffs.

[sanman]

Look Ma, No Wires!

Tesla's Optimus robot has had an update which shows it dancing in a very human-like way:




Musk and the Tesla people say these abilities will extend to performance improvements well beyond the field of dance itself.

How long until table tennis?

[sanman]

"Freeze All Motor Functions! I said Freeze All Motor Functions!" 

https://twitter.com/Tesla_Optimus/status/1922456791549427867

[Twark_Main]

"Freeze All Motor Functions! I said Freeze All Motor Functions!" 

...

Those who danced were thought to be quite insane by those who couldn't hear the music.

The most recently leaked footage from the H1 incident is shedding new light on the root cause investigation. See attached


We've all been there, haven't we?

[Robotbeat]

"Freeze All Motor Functions! I said Freeze All Motor Functions!" 

...

Those who danced were thought to be quite insane by those who couldn't hear the music.

The most recently leaked footage from the H1 incident is shedding new light on the root cause investigation. See attached


We've all been there, haven't we?

wat

[Twark_Main]

"Freeze All Motor Functions! I said Freeze All Motor Functions!" 

...

Those who danced were thought to be quite insane by those who couldn't hear the music.

The most recently leaked footage from the H1 incident is shedding new light on the root cause investigation. See attached


We've all been there, haven't we?

wat

I expect investigators recovered the files uploaded to the bot and synchronized them to video. In future testing no bass drops harder than C+ tier until the root cause of "Going Too Hard" can be addressed.   

[sanman]

Can there one day be a Moore's Law type of performance metric based on dance moves?

Notice how it's adjusting its footing on the fly, while letting its feet slip.

Will a machine like this one day do a Michael Jackson moonwalk on our actual Moon?
(Make it happen, Elon!)

Its active stabilization/control in combination with its natural instability would mean that it can never get stuck in the sand like the simpler 6-wheeled Spirit or Opportunity.

[daedalus1]

What is this?

Someone told Optimus to do a particular dance.

It was a preprogrammed sequence.

Someone was wearing sensors and the robot copied their dance move.

If it's not the first one then it's not really significant.

[Cheapchips]

What is this?

Someone told Optimus to do a particular dance.

It was a preprogrammed sequence.

Someone was wearing sensors and the robot copied their dance move.

If it's not the first one then it's not really significant.

Tesla aren't the first, since we've seen Unitree, Boston etc do similarly complex full body motion. What is significant is the 'how it's done' of all these recent demos.

You can't just map full body mocap of any action and then have a robot perform it.  Even with inverse kinematics, physics takes over and you end up with heap of robot on the floor.

With reinforcement learning the mocap or animation data is fed into a simulation where the simulated robot tries to perform that action hundred of millions of times.  Natural selection from failure eventually creates a policy which can execute the moves successfully across a range of scenarios, including the natural variability in a robots own actuation. It's not an animation playback.

Tesla and others are currently keen to point out that you RL train the action, drop the policy into the robot and it just works.

The great white hope is that this same methodology works for complex full body manipulation. The real killer demo is going to be a robot picking something up in one hand and opening a door with the other.

[sanman]

Tesla aren't the first, since we've seen Unitree, Boston etc do similarly complex full body motion. What is significant is the 'how it's done' of all these recent demos.

You can't just map full body mocap of any action and then have a robot perform it.  Even with inverse kinematics, physics takes over and you end up with heap of robot on the floor.

With reinforcement learning the mocap or animation data is fed into a simulation where the simulated robot tries to perform that action hundred of millions of times.  Natural selection from failure eventually creates a policy which can execute the moves successfully across a range of scenarios, including the natural variability in a robots own actuation. It's not an animation playback.

Tesla and others are currently keen to point out that you RL train the action, drop the policy into the robot and it just works.

The great white hope is that this same methodology works for complex full body manipulation. The real killer demo is going to be a robot picking something up in one hand and opening a door with the other.

I'd imagine a major killer demo could be in robot athletic sports.

Set up a Robot Sports League that makes use of standardized robots which they supply, and then have competing sports teams drop their respective policies into them, to compete on that basis. And since AI's can support multiple policies simultaneously coexisting together, then the Robot Sports League could also supply the policies for adhering to the rules of a given particular sport.

Teams could then compete for recognition, prize money, sponsorships, etc among their various incentives, just as happens with classical professional sports.