TechDaily.ai

techdaily.ai, your source for technical information. This deep dive is sponsored by Stonefly, your trusted solution provider and adviser in enterprise storage, backup, disaster recovery, hypercon converged and VMware, HyperV, Proxmox cluster, AI servers, SA365, a turnkey security appliance, and public and private cloud. Check out stonefly.com or email your project requirements to sales@ stonefly.com. Okay, let's dive right into some something really fascinating today.
Embodied AI.
We're talking about uh moving AI beyond just the code, the algorithms, and actually putting it inside physical machines,
right? Giving it a body.
Exactly. Think less, you know, brain in a jar and more like a complete organism, something that actually interacts with the world around it.
And the intelligence itself comes from that interaction.
That's the key point, isn't it? It's the physicality, the senses, the environment, they're all inseparable from the intelligence.
It's a fundamental shift from how we've traditionally thought about AI.
So today, we're really exploring this uh this intersection of the physical and the digital.
Yeah, it's a really exciting space. Traditional AI, well, it often saw intelligence as pure computation, right? Abstract, right?
Embodied AI says no, having a body, sensing the world through that body, acting with that body. That's not just an add-on. It fundamentally shapes intelligence.
It changes the whole game for how you'd create and I guess test these systems
completely.
And when you talk about giving AI a physical form, especially one that maybe looks and acts like a living thing. This idea of physical AI. Yeah.
Well, that sounds like a huge leap.
It is a massive one.
What kind of skills do you even need for that? It feels like it must pull from everywhere.
Oh, it's incredibly interdicciplinary. I mean, think about it. To make a robot that moves and senses like say an animal or a person.
You need materials, science, definite
for the body itself,
like flexible materials.
Exactly. New polymers, composits, things with specific properties. Then, uh mechanical engineering is core robotics. Obviously mechatronics
that's the mix of mechanical and electronic
right plus manufacturing design and then the computer science side is crucial
programming ML data science
all of that but then you also need biology understanding physiology maybe even tissue engineering biomechanics how things actually move
wow even chemistry
yeah absolutely synthesizing materials understanding biochemical interactions potentially it's this uh this real fusion of fields
not just code anymore. It's getting these very different disciplines to talk to each other.
Precisely. And that integration is the hard part.
So, this brings me to the big question. How do you measure progress? How do you know if this embodied AI is actually getting smarter in a meaningful way? What are the tests?
That's uh that's the core challenge, isn't it? Our old ways of testing AI, well, they start to look a bit limited.
Like the touring test.
Exactly. The original touring test from what 1950 was all about language. Can a machine fool a human judge? just through text conversation
and there was that 30% benchmark, right? Some chat bots claim to pass.
Yeah. Like Eugene Gooseman. But those passes are often debated, you know, depending on the setup, the AI's persona,
right? It feels very linguistic.
Yeah.
But for a robot interacting physically, talking is just one small part.
Precisely. How do you test the doing? This leads to the idea of a physical touring test.
Okay. What does that look like?
Well, imagine you come home, your apartment is spotless. Dinner's perfectly cooked and you genuinely cannot tell if a human or a machine did it all.
Ah, so it's about indistinguishable physical task performance.
Exactly. Complex realworld tasks perform seamlessly.
That sounds incredibly difficult. I mean, current robots still struggle with seemingly simple things sometimes, like picking up unfamiliar objects reliably.
Oh, absolutely. We're a long way off from that vision. The fine motor control, the adaptability, it's just not there yet.
What's the main hurdle?
A huge one is The data challenge in robotics, it's quite different from language models.
How so?
Well, language models learn from, you know, the vast ocean of text and images on the internet. But for robots, to learn smooth, continuous joint movements for complex actions,
you need specific examples,
very specific, often requiring a human to physically demonstrate the task, sometimes over and over. Recording that continuous control data is painstaking.
So, you just don't have the sheer volume of training data you have for language.
Not even close. It's much scarcer and harder to get.
Okay, so the original touring test isn't enough and the physical touring test seems like a very distant goal. Are there intermediate steps? Other ways to evaluate these more lifelike robots?
Yes, thankfully there's a concept called the multimodal touring test or MTT. It's designed to be more uh holistic for these realistic humanoid robots.
Look at modal. So it looks at different aspects,
right? It breaks down human likeness into testable parts. Appearance is one. Does it look visually indistinguishable.
Okay.
Then movement, how authentic and synchronized are its actions? Voice is another natural speech. And crucially, how the mouth moves with the sound.
Lip syncing basically
essentially. Yes. And finally, the AI itself, its cognitive abilities may be tested through interactive games or tasks.
That sounds much more comprehensive breaking it down like that.
Mhm. And the MT also proposes stages sort of like developmental milestones.
Whoa. Like what?
Well, there's fragmentaryary ation. Maybe just the hand movements look perfectly human. Then synchronized emulation where say movement and voice work together seamlessly.
And the final goal
absolute emulation where everything comes together appearance, movement, voice, AI making it truly indistinguishable.
Okay, that feels like a more practical road map. You can see progress along the way, not just past fail.
Exactly. It provides a framework for development.
This also reminds me our own intelligence isn't purely mental, is it? It's tied to our bodies.
This idea A of embodied cognition.
You've got it. Embodied cognition is key here. It emphasizes how our physical bodies are crucial for efficient thinking. The brain doesn't do all the work alone.
It offloads some processing.
Sort of. Yeah. It leverages the body and its interaction with the environment. Think about walking, swinging your arms, helps with balance, right? Your brain isn't consciously calculating every tiny adjustment. A lot of that stability comes from the physical mechanics, the interaction with the ground.
That makes sense. So, if Intelligence is so tied to the body. What does that mean for testing embodied AI? Simulations maybe aren't enough.
That's a really critical point. Simulations are useful, no doubt, especially early on,
but they're not the real world.
Never completely because these systems are designed for the physical world. You ultimately have to test them there in realistic contexts. Simulations will always miss some complexity, some unexpected factor.
You can't know how it really performs until it's out there. pretty much. And there's another layer to this testing challenge. The nature of neural networks,
how they learn
and how they store knowledge. Unlike old rule-based AI where you could see the logic, in neural nets, knowledge is kind of smeared across millions of connections, it's implicit,
making it harder to understand why it does something like a black box.
It can be. Yes. This opacity makes debugging and thorough testing harder. It's a bit like testing a trained animal versus testing a machine with clearance. instructions. You see the behavior, but the internal reasoning isn't always transparent.
You test the performance in the real world because that's what ultimately matters.
Exactly. Real world performance is the benchmark.
This has been well really eye opening. Moving AI into physical bodies. It clearly opens up huge possibilities. More capable, adaptable machines interacting with our world.
The potential is enormous.
But it also throws up massive challenges getting different experts to work together. finding enough data and figuring out these new ways to actually test them like the physical and multimodal touring tests.
Absolutely. Bridging that gap between the digital and physical requires these interdisciplinary approaches and new ways of evaluating progress. We need those frameworks like the MTT.
It also makes you think, are there ethical considerations here even if these machines aren't conscious in the way we are? I remember reading something about potential non-concious cognitive suffering.
That's definitely an emerging area of ethical discussion as these systems become more complex and interact physically, even without consciousness. Questions about their treatment and potential negative states might arise. It's something we need to keep in mind as the technology advances.
Right? So, here's a a final thought maybe for you, the listener, to ponder as we create machines that sense and act more like us, interacting with the physical world. What completely new ways will we need to invent just to understand them, to evaluate what they can do and uh what their impact might be?
It feels like we're just scratching the surface.
Yeah, he really does. A whole new frontier. techaily.ai, your source for technical information. This deep dive is sponsored by Stonefly, your trusted solution provider and adviser in enterprise storage, backup disaster recovery, hypercon converged in VMware, HyperV, Proxm cluster, AI servers, SA365, a turnkey security appliance, and public and private cloud. Check out stonefly.com or email your project requirements to salesstonefly.com.