Humanoid Robots for Moon and Mars

[Coastal Ron]

Also, computer technology isn't advancing that quickly. GPU's are, but CPU's certainly are not, and CPU's are critical components for running the operating systems of robots.

AI is making fast advances, but that only applies to part of the software needed for bi-pedal robots, and the hardware for robots is not advancing that fast, and is still pretty rudimentary compared to what a human can do.

The hardware will evolve to look more like our brain, where the processing and memory are more co-located / co-resident.
Our brain has memory and processing more distributed and intertwined with each other.

Recently I was reading an article that advocated that computer "brains" would NOT evolve to mimic human brains, since human brains evolved because of our unique organic needs, but computer brains don't need to follow (and probably can't) that method of evolving.

Regardless, look at the velocity of progress overall, over the past decade, and you'll see that in the real world we experience when we walk around, AI systems have not made much of a difference. Robotic systems even less. Certainly my 2021 Tesla Model Y hasn't evolved to be any smarter, and people buying a 2025 Tesla Model Y don't get much more than what I experience. Explain that.

We can see that GPUs, whether from Nvidia or competitors like AMD, are accruing more short-term cache memory over time... This is clearly a harbinger for the future of where things are going.

I have never advocated that progress was not being made. I've lived through Moore's Law, which many think may be reaching its natural end (i.e. you can't shrink atoms), but that just means humanity will discover some other method of evolving computing systems. The question I'm debating is HOW QUICKLY computing systems will evolve that can control humanoid robotic systems, and based on what can be seen - not predicted, but seen - they are not yet evolving very quickly.

Which is why I think human-assisted or tele-robotic systems will be of the most use for space exploration.

For the near future. But for the longer-term future, autonomy will naturally be more useful and advantageous. AI will allow our tele-robotic commands to be given at more of a high-level for execution rather than having to be low-level control and micro-management.

The question is "when". Next year, next decade? When?

I'm not one to believe PR hype, especially since so many "experts" such as Elon Musk have been so wrong with their predictions of what will happen with AI. If you had a friend that was always over exaggerating, you would tend to discount anything they predicted, right? That is where I am at with the whole realm of so-called "AI experts", and why I'm now in the mode of "show me, don't tell me". 

[TheRadicalModerate]

Recently I was reading an article that advocated that computer "brains" would NOT evolve to mimic human brains, since human brains evolved because of our unique organic needs, but computer brains don't need to follow (and probably can't) that method of evolving.

There's a significant counterexample that casts this statement into doubt:  Tesla recently transitioned the FSD suite from a rules-based framework that used some auxiliary neural network modules, to a neural-network-based framework that uses some auxiliary rules.  IOW, they're becoming more brain-like, not less.

We can obviously have a long discussion where we bicker about what it means to be "brain-like"; let's not.  But the key distinction between rules-based systems and neural networks is important.

Given Tesla's current software architecture proclivities, I'd bet that Optimus will also become more neural-network-based.

What would this mean for hybridizing teleoperation and true autonomous operation?

I can imagine a humanoid robot that starts out with a basic set of rough-surface locomotion skills, then trains itself on specific fine-motor tasks, based on the experience provided by the teleoperators performing those tasks.  When it's confident enough in its actions, the tasks can be offloaded from the teleoperators and proceed autonomously.

That's a big step from where we are now, but it's a step that is consistent with how machine learning is changing.

Update:  Here's a Boston Dynamics blog post about using reinforcement learning with Spot.

[InterestedEngineer]

Recently I was reading an article that advocated that computer "brains" would NOT evolve to mimic human brains, since human brains evolved because of our unique organic needs, but computer brains don't need to follow (and probably can't) that method of evolving.

There's a significant counterexample that casts this statement into doubt:  Tesla recently transitioned the FSD suite from a rules-based framework that used some auxiliary neural network modules, to a neural-network-based framework that uses some auxiliary rules.  IOW, they're becoming more brain-like, not less.

We can obviously have a long discussion where we bicker about what it means to be "brain-like"; let's not.  But the key distinction between rules-based systems and neural networks is important.

Given Tesla's current software architecture proclivities, I'd bet that Optimus will also become more neural-network-based.

What would this mean for hybridizing teleoperation and true autonomous operation?

I can imagine a humanoid robot that starts out with a basic set of rough-surface locomotion skills, then trains itself on specific fine-motor tasks, based on the experience provided by the teleoperators performing those tasks.  When it's confident enough in its actions, the tasks can be offloaded from the teleoperators and proceed autonomously.

That's a big step from where we are now, but it's a step that is consistent with how machine learning is changing.

Update:  Here's a Boston Dynamics blog post about using reinforcement learning with Spot.

As a surfer I climb over slippery slime covered rocks all the time.

I find it amusing that they haven't programmed the Boston Dynamics robot to reach out a foot and test the slipperiness of the next step.  I do it all the time, whether it's a foot or the one free hand (or my eyes which can tell the difference between brown slime and a barnacle covered rock, which is not slippery at all but you need booties to step on).

So I'm curious whether they will use a rules based system to eventually get that behavior, or whether the AI will learn it.

If they don't have any cold water surfers or jetty fishermen on staff they might never figure out it's a useful strategy.

But will their AI have enough degrees of freedom to figure it out on their own?

[edzieba]

The Neural Networks being run at scale today are VERY, VERY far from being 'brain-like' at the architectural level, let alone at the hardware level. There have been many attempts at brain-like physical chip architectures for neural network computation, and every single one has fallen flat on its face as being dramatically slower and less efficient compared to running on serialised processors. In Memory Computation has not fared much better when it comes to actual implementations, as by the time sufficient operations are built into the IMC die to actually be useful for NNs what you end up with is a large matrix acceleration unit with a really terrible memory architecture that is actively hostile to cross-chip memory access or to initial data ingest (i.e. any speedups from having more local memory are more than lost from introduced delays in getting data in and out).
Even at a fundamental level, brains are timing-based systems operating asynchronously and modulating mostly in the frequency domain (pulse-train modulation). All current practical ANN implementations are synchronous clocked with value-based modulation. ANNs are almost entirely connectome-free (barring some partitioning in the case of clustered training which is minimised to the absolute maximum extent possible) whereas actual brains not only have a connectome but it is very strongly correlated with physical neuron position.

[Coastal Ron]

Recently I was reading an article that advocated that computer "brains" would NOT evolve to mimic human brains, since human brains evolved because of our unique organic needs, but computer brains don't need to follow (and probably can't) that method of evolving.

There's a significant counterexample that casts this statement into doubt:  Tesla recently transitioned the FSD suite from a rules-based framework that used some auxiliary neural network modules, to a neural-network-based framework that uses some auxiliary rules.  IOW, they're becoming more brain-like, not less.

Unless Tesla has implemented some radical new chip architectures into Hardware 4 (AI4) vs Hardware 3, it is just an extension of what already exists, but with a different approach to the software.

We can obviously have a long discussion where we bicker about what it means to be "brain-like"; let's not.  But the key distinction between rules-based systems and neural networks is important.

Right, let's not, especially since humanity is still trying to figure out how human brains work. Suffice it to say that Tesla is still iterating on the same technology path that everyone else is using, with just different software and data.

Given Tesla's current software architecture proclivities, I'd bet that Optimus will also become more neural-network-based.

Maybe. However I'm not sure the software needs of a humanoid robot are similar enough to a car, meaning that they might not get much help from the car hardware.

I can imagine a humanoid robot that starts out with a basic set of rough-surface locomotion skills, then trains itself on specific fine-motor tasks, based on the experience provided by the teleoperators performing those tasks.  When it's confident enough in its actions, the tasks can be offloaded from the teleoperators and proceed autonomously.

Watch, learn, do. Or something to that effect.

That's a big step from where we are now, but it's a step that is consistent with how machine learning is changing.

Sure, but as I said in the post above:

The question I'm debating is HOW QUICKLY computing systems will evolve that can control humanoid robotic systems, and based on what can be seen - not predicted, but seen - they are not yet evolving very quickly.

Plenty of hype, but the velocity of change is still slow in my opinion. However, for doing work on Mars, if you have the mass budget to send them there, then absolutely we should try out a variety of humanoid robots. Especially since there are only some things we can learn (and the robots can learn) by actually being there. So experiment, just don't bet the farm (or colony) yet on humanoid robots doing everything.

[sanman]

The Neural Networks being run at scale today are VERY, VERY far from being 'brain-like' at the architectural level, let alone at the hardware level. There have been many attempts at brain-like physical chip architectures for neural network computation, and every single one has fallen flat on its face as being dramatically slower and less efficient compared to running on serialised processors. In Memory Computation has not fared much better when it comes to actual implementations, as by the time sufficient operations are built into the IMC die to actually be useful for NNs what you end up with is a large matrix acceleration unit with a really terrible memory architecture that is actively hostile to cross-chip memory access or to initial data ingest (i.e. any speedups from having more local memory are more than lost from introduced delays in getting data in and out).
Even at a fundamental level, brains are timing-based systems operating asynchronously and modulating mostly in the frequency domain (pulse-train modulation). All current practical ANN implementations are synchronous clocked with value-based modulation. ANNs are almost entirely connectome-free (barring some partitioning in the case of clustered training which is minimised to the absolute maximum extent possible) whereas actual brains not only have a connectome but it is very strongly correlated with physical neuron position.

So what's wrong with operating asynchronously? Why is it absolutely necessary to stay on a common clock? There's nothing wrong with moving toward that architecture. Intel's Itanium concept may have to make a comeback (analogous to the way Musk brought back some of the Soviet N-1 design concepts and made them work - what's old is new again)
Like a free-market economy full of people each all doing their own thing independently, instead of functioning like a command economy where everybody is marching along in lockstep, all taking their bathroom breaks at the same time, computing could likewise benefit from more diversified processing with variable timing. That's not merely a biological thing, that's a physics thing.

[InterestedEngineer]

Yall are thinking at the circuit level.  CMOS circuit design is far different than wetware, it's going to be different architecture.

Think instead at the macro level.  The human brain isn't one big Neural Net.  It's a bunch of smaller specialized NN cooperating.

There's specialization for hearing, seeing, emotions, higher coordination functions, lower coordination functions, etc.  Neuroscientists have mapped this all out over the last 100+ years.

Some of those brain functions may be more built in logic (expert systems) than NN.

I have yet to see the "cooperating neural net" approach taken on various systems.  It's mostly just centralized NNs.

The closest I've seen is training a NN to take over some expert system's function (per the article above).

[sanman]

Here's some company/team that are claiming a different approach (how different it actually is, I can't really tell yet)




But just wanting to address previous points made upthread about not having to resemble biology.
Look, a computer mouse has a shape that correlates to your hand . Why? Because our hand is going to use it.
A humanoid robot has a shape that correlates to humans like you and I. Why? Because it's going to be navigating and maneuvering in environments already designed for us to comfortably function in.
If we're talking about computers and robots designed to handle tasks that you and I have traditionally done, then what's wrong with having AI that resembles how our own brains work? That doesn't have to be inherently confining - it all depends on what we're trying to accomplish.

[catdlr]

Robots that can operate cameras on other planets?

Humanoid Robot Atlas Takes on a New Role as a Camera Operator

Apr 2, 2025
Boston Dynamics’ Atlas is taking on a new role—this time, behind the camera. In a collaboration with WPP, Canon, and NVIDIA, Atlas has been tested as a robotic camera operator, demonstrating how humanoid robots could assist in filmmaking. With its ability to hold heavy equipment, execute precise movements, and train in virtual environments, Atlas represents a potential shift in how robotics integrate into creative and industrial fields.

[Twark_Main]

Yall are thinking at the circuit level.  CMOS circuit design is far different than wetware, it's going to be different architecture.

Think instead at the macro level.  The human brain isn't one big Neural Net.  It's a bunch of smaller specialized NN cooperating.

There's specialization for hearing, seeing, emotions, higher coordination functions, lower coordination functions, etc.  Neuroscientists have mapped this all out over the last 100+ years.

Some of those brain functions may be more built in logic (expert systems) than NN.

I have yet to see the "cooperating neural net" approach taken on various systems.  It's mostly just centralized NNs.

The closest I've seen is training a NN to take over some expert system's function (per the article above).

In practice, "cooperating neural nets" are equivalent to just weird-shaped, multi-headed neural nets (or, worse, separate NNs glued together by hand-written code ala Autopilot 1.0).

The search term is "Hydranet" or "multi-headed model."

https://stackoverflow.com/questions/56004483/what-is-a-multi-headed-model-and-what-exactly-is-a-head-in-a-model

[StraumliBlight]

https://twitter.com/GITAI_HQ/status/1907387602249957700

Revolutionizing Lunar Operations: Robots at Work 🌕🤖
Open Positions ➡️ https://job-boards.greenhouse.io/gitai

Check out our #GITAI Lunar Rover autonomously assembling lunar infrastructure and deploying solar panels in a simulated lunar environment.

https://twitter.com/GITAI_HQ/status/1905213247944372275

[InterestedEngineer]

https://twitter.com/GITAI_HQ/status/1907387602249957700

Revolutionizing Lunar Operations: Robots at Work 🌕🤖
Open Positions ➡️ https://job-boards.greenhouse.io/gitai

Check out our #GITAI Lunar Rover autonomously assembling lunar infrastructure and deploying solar panels in a simulated lunar environment.

https://twitter.com/GITAI_HQ/status/1905213247944372275

is that remote control or training?

[Roy_H]

Here's some company/team that are claiming a different approach (how different it actually is, I can't really tell yet)

What the video fails to appreciate, is that much of AI development is deliberately avoiding the memory and free learning proposed. This is because of concerns over loosing control of the AI or robot. It will come, it is just a frightening process to build a robot with an AI much smarter than we are without any controls about what it will do. Will it help us or turn on us?

[sanman]

So not relevant to this thread then.

Do we really need a separate robot-horse thread, though?

This concept could be useful on the irregular surfaces of the Moon and Mars, for navigating where dune-buggies can't.

Autonomously it could still be used as a mule to carry payloads.

[daedalus1]

So not relevant to this thread then.

Do we really need a separate robot-horse thread, though?

This concept could be useful on the irregular surfaces of the Moon and Mars, for navigating where dune-buggies can't.

Autonomously it could still be used as a mule to carry payloads.

Yes. The title is quite specific.

[lamontagne]

The Gitai robot is interesting but is at advantageous compared to an Optimus riding in a Cybertruck solution (both significantly adapted)?  Perhaps in a mass constrained mission, but less so on a Starship type mission?

[sanman]

Here's a more anatomically accurate humanoid robot which more closely imitates the human skeleto-musculature





Maybe something like this could put in a spacesuit, to better stress-test the thing across many hours of human-like movements.

[clongton]

Here's a more anatomically accurate humanoid robot which more closely imitates the human skeleto-musculature





Maybe something like this could put in a spacesuit, to better stress-test the thing across many hours of human-like movements.

Visions of Mr. Data from Startrek TNG, or Cylons from Battlestar Galactica, or the Borg.

[Twark_Main]

Here's a more anatomically accurate humanoid robot which more closely imitates the human skeleto-musculature


-


Maybe something like this could put in a spacesuit, to better stress-test the thing across many hours of human-like movements.

Visions of Mr. Data from Startrek TNG, or Cylons from Battlestar Galactica, or the Borg.

Hey, I actually liked the classic TNG episode "Mr. Data Dons and Doffs A Standard Issue Starfleet Pressure Suit Nine Hundred Times." Bold directoral move to not cut away on the 673rd doffing!   

[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.