It does all start to feel like we'd get fairly close to being able to convincingly emulate a lot of human or at least animal behavior on top of the existing generative stack, by using brain-like orchestration patterns ... if only inference was fast enough to do much more of it.
The gauge-reading example here is great, but in reality of course having the system synthesize that Python script, run the CV tasks, come back with the answer etc. is currently quite slow.
Once things go much faster, you can also start to use image generation to have models extrapolate possible futures from photos they take, and then describe them back to themselves and make decisions based on that, loops like this. I think the assumption is that our brains do similar things unconsciously, before we integrate into our conscious conception of mind.
I'm really curious what things we could build if we had 100x or 1000x inference throughput.
Idk if you've seen this already but Taalas does this interesting thing where they embed the model directly onto the chip, this leads to super-fast speeds (https://chatjimmy.ai) but the model they're using is an old small Llama model so the quality is pretty bad. But they say that it can scale, so if that's really true that'd be pretty insane and unlock the inference you're talking about.
Robotics/control systems is exactly what came to mind when I saw this release! What struck me is the possibility of look ahead search in real time, a bit like alphazero's mcts.
Actually pretty interesting to think: in a few years you might buy a raspberry pi style computer board with an extra chip on it with one of these types of embodiment models and you can slap it in a rover or something.
Is emulating human behavior really a valuable end goal though? Humans exist as the evolutionary endpoint of exhaustion hunting large pray and organic tool-making. We've built loads of industrial and residential automation tools in the last 100 years and none of them are humanoid. I'd imagine a household robot butler would be more like R2D2 with lots and lots of arms.
It is when the world was made to interface with us. We can't use robots for everything if they aren't emulating us, because we would have to adapt everything for the non-humanlike robots.
Every single behavior? For sure not but otherwise we are the result of a very very long evolution and there is nothing else around us as smart and as adjustable.
The planing ahead thing through simulation for example seems to be a very good tool in neuronal network based architectures.
Maybe dumb question: One of the use cases is instrument reading of analog instruments. My brain immediately goes to "this should have some sensor sending data, and not be analog". Is having a robot dog read analog sensors really a better fit in some cases?
What if it turns out that "take care of this land" means the traditional way California was taken care of with regular small slow burns. After over 10k years of this type of management there are many important native species that won't even germinate without the presence of ash.
Or it could turn out to look like satayoma (Japanese peasant forests) or it could be more similar to the crop rotation that was traditionally practiced in many parts of Central Africa where roots were important.
In Russia before the Soviets forced "modern scientific agriculture" on peasants to modernize, they practiced things like contour farming (where they interplanted rows of crops against the contours of the land to slow water down) and maslins (where they intermixed multiple varieties of wheat and barleys in the same patch). Now contour farming are an active area of research for their ability to prevent topsoil loss and build soil health while maslins provide superior yield stability and use little to no pesticides.
That's not even getting into the over 40-120,000 varieties of rice we've documented. Most of which are hyper adapted to a very specific location—often even a single village.
My point is there is no one way to take care of a plot of land. It's all relative to a number of factors beyond just the abiotic characteristics of the land itself. Your goals and intentions matter and you will always find localized unique adaptations.
Are you saying that has failed? It isn't obvious to me from that page that anything in particular is going wrong. I don't think anyone is daft enough to claim that AI solves the "Iowa remains unplantable due to winter conditions" problem.
logs suggest it's been 'critically failing' and 'blocked for 68 days' on farmhand introduction, although the logs don't go back far enough (and cut off too early) to really tell what's going on. https://proofofcorn.com/log
That's kind of the opposite problem -- the agent doesn't have robot arms or legs or a parcel of land. It has to rely on people to get access to land and plant and harvest the corn, and those people are ignoring it.
>Our safest robotics model yet
Safety is integrated into every level of our embodied reasoning models. Gemini Robotics-ER 1.6 is our safest robotics model to date, demonstrating superior compliance with Gemini safety policies on adversarial spatial reasoning tasks compared to all previous generations.
The safety guidelines are interesting, they treat them as a goal that they are aspiring to achieve, which seems realistic. It’s not quite ready for prime time yet.
Pointing a camera at a pressure gauge and recording a graph is something that I would have found useful and have thought about writing. Does software like that exist that’s available to consumers?
This seems perfect to hook up to my 'LLMs can control robots over MCP' system. The idea is that LLMs are great at writing code, so let's lean in to that. I'll give it a try! I just got a bigger robot, we'll see how it does...
Really unfortunate that I forgot what YT video i saw about this just 2 weeks ago.
It was about Googles PaLM-E evolution and progress. It basically has two models one which controls the robot, the other is a llm and they are combined together in some attention layer.
That video is pretty good. I'm basically betting that an earlier, abandoned approach described in the video, "Code as Policy", will beat everything else. It requires no training data, and generalizes instantly to all robots.
I’ve been thinking about AI robotics lately… if internally at labs they have a GPT-2, GPT-3 “equivalent” for robotics, you can’t really release that. If a robot unloading your dishwasher breaks one of your dishes once, this is a massive failure.
So there might be awesome progress behind the scenes, just not ready for the general public.
> If a robot unloading your dishwasher breaks one of your dishes once, this is a massive failure.
That's a bit exaggerated, no? Early roombas would get tangled in socks, drag pet poop all over the floor, break glass stuff and so on, and yet the market accepted that, evolved, and now we have plenty of cleaning robots from various companies, including cheap spying ones from china.
I actually think that there's a lot of value in being the first to deploy bots into homes, even if they aren't perfect. The amount of data you'd collect is invaluable, and by the looks of it, can't be synth generated in a lab.
I think the "safer" option is still the "bring them to factories first, offices next and homes last", but anyway I'm sure someone will jump straight to home deployments.
I ended up watching Bicentennial Man (1999) with Robin Williams over the weekend. If you haven't seen I thought it was a good and timely thing to watch and is kid friendly. Without giving away the plot, the scene where it was unloading the dishwasher...take my money!
From an economic standpoint the industry is anyway the most relevant by far. Its easier as the env is a lot more controlled, professionals configure and maintain the robots, they buy in bulk and have more money.
My concern with a household robot is not the dishwasher but the tv screen, the glas door, glas table, animals (fish/aquarium) etc. the robot might walk through, touch through or fall onto.
If someone paid 100 grand for you to load and unload the the dishwasher, and the research to be able to do it costed hundreds of billions, decades of research, hundreds of thousands of researchers, and that was the ONLY thing you could do, yes, you WOULD be a massive failure.
> If a robot unloading your dishwasher breaks one of your dishes once, this is a massive failure.
Depending on what the rate of breaking dishes is, this would be a massive improvement on me, a human being, since I break a really important dish I needed to use like ~2x per month on average.
VLA models essentially take a webcam screenshot + some text (think "put the red block in the right box") and output motor control instructions to achieve that.
Note: "Gemini Robotics-ER" is not a VLA, though Gemini does have a VLA model too: "Gemini Robotics".
Meanwhile, gemini 3.1 pro (that was released two months ago) was completely unavailable to me this afternoon, neither with API nor Subscription.
Nothing was reported in Google status page, not even the CLI is responding, it’s just left there waiting for an answer that will never arrive even after 10 minutes.
Nice. I couldn't find the part that I'm most interested in though, latency. This beats their frontier vision model for some identification tasks -- for a robotics model, I'm interested in hz. Since this is an "Embodied Reasoning" model, I'm assuming it's fairly slow - it's designed to match with on-robot faster cycle models.
In my quick image recognition testing on AI studio, it's performance seems similar to 3.1 pro, but is much much faster. It "thinks" but only for a few seconds.
Of course this is for counting animal legs while giving coordinates and reading analog clocks. Not coding or or solving puzzles. I imagine the image performance to model weight of this model is very high.
I thought the entire discourse on how important pointing is to be super interesting. I've been told, although I don't know if this is true, that dogs are the only animal that can understand human pointing. Fascinating to think this might be fundamental to world intelligence requirements. Well, it's required, but it's interesting to think that it might be a core structure required or that learning it might force some sort of neural architecture that's helpful.
And, I was disappointed to see that pointing was just giving x,y coords. I wanted to see robots pointing at stuff.
Would this approach destroy critical investments in physics- or modeling-based reasoning?
I'm all for the task reasoning and the multi-view recognition, based on relevant points. I'm very uncomfortable with the loose world "understanding".
The fault model I see is that e.g., this "visual understanding" will get things mostly right: enough to build and even deliver products. However, these are only probabilistic guarantees based on training sets, and those are unlikely to survive contact with a complex interactive world, particularly since robots are often repurposed as tasks change.
So it's a kind of moral-product-hazard: it delivers initial results but delays risk to later, so product developers will have incentives to build and leave users holding the bag. (Indeed: users are responsible for integration risks anyway.)
It hacks our assumptions: we think that you can take an MVP and productize it, but in this case, you'll never backfit the model to conform to the physics in a reliable way. I doubt there's any way to harness Gemini to depend on a physics model, so we'll end up with mostly-working sunk investments out in the market - slop robots so cheap that tight ones can't survive.
I was also just in the market for a small experiment robot. I got the hiwonder armpi-fpv. Avoid it, the actuators are pretty bad - they're very 'grindy', the robot jitters like crazy when it moves. Any such problems with the lekivi?
As the article notes regular Gemini and Gemma also have spatial reasoning capabilities, which I decided to test by seeing if Gemini could drive a little rover successfully, which it mostly did: https://martin.drashkov.com/2026/02/letting-gemini-drive-my-...
LLMs are really good at the sort of tasks that have been missing from robotics: understanding, reasoning, planning etc, so we'll likely see much more use of them in various robotics applications. I guess the main question right now is:
- who sends in the various fine-motor commands. The answer most labs/researchers have is "a smaller diffusion model", so the LLM acts as a planner, then a smaller faster diffusion model controls the actual motors. I suspect in many cases you can get away with the equivalent of a tool call - the LLM simply calls out a particular subroutine, like "go forward 1m" or "tilt camera right"
- what do you do about memory? All the models are either purely reactive or take a very small slice of history and use that as part of the input, so they all need some type of memory/state management system to actually allow them to work on a task for more than a little while. It's not clear to me whether this will be standardized and become part of models themselves, or everyone will just do their own thing.
For the fine-motor commands: or, the model can write the code to generate them on the fly. It seems to work, in my very limited experiment.
As for memory: my approach is to give the robot a python repl and, basically, a file system - the LLM can write modules, poke at the robot via interactive python, etc.
Basically, the LLM becomes a robot programmer, writing code in real-time.
Showing the murder dog reading a gauge using $$$ worth of model time is kinda not an amazing demo. We already know how to read gauges with machine vision. We also know how to order digital gauges out of industrial catalogs for under $50.
I think that where this gets interesting is when you can just drop these robotic systems into an environment that wasn't necessarily set up specifically to handle them. The $50 for your gauge isn't really the cost: it's engineering time to go through the whole environment and set it up so that the robotic system can deal with each of the specific tasks, each of which will require some bespoke setup.
Agree. I'm unclear what's the highlight of this post. Is the multimodality of the model (that can replace computer vision), is it the reasoning part, is it the overall wrapper that makes it very easy to develop on top?
It does all start to feel like we'd get fairly close to being able to convincingly emulate a lot of human or at least animal behavior on top of the existing generative stack, by using brain-like orchestration patterns ... if only inference was fast enough to do much more of it.
The gauge-reading example here is great, but in reality of course having the system synthesize that Python script, run the CV tasks, come back with the answer etc. is currently quite slow.
Once things go much faster, you can also start to use image generation to have models extrapolate possible futures from photos they take, and then describe them back to themselves and make decisions based on that, loops like this. I think the assumption is that our brains do similar things unconsciously, before we integrate into our conscious conception of mind.
I'm really curious what things we could build if we had 100x or 1000x inference throughput.
Idk if you've seen this already but Taalas does this interesting thing where they embed the model directly onto the chip, this leads to super-fast speeds (https://chatjimmy.ai) but the model they're using is an old small Llama model so the quality is pretty bad. But they say that it can scale, so if that's really true that'd be pretty insane and unlock the inference you're talking about.
Robotics/control systems is exactly what came to mind when I saw this release! What struck me is the possibility of look ahead search in real time, a bit like alphazero's mcts.
Taalas showed that you could make LLMs faster by turning them into ASICs and get 10k+ token generation. It's a matter of time now.
Actually pretty interesting to think: in a few years you might buy a raspberry pi style computer board with an extra chip on it with one of these types of embodiment models and you can slap it in a rover or something.
Is emulating human behavior really a valuable end goal though? Humans exist as the evolutionary endpoint of exhaustion hunting large pray and organic tool-making. We've built loads of industrial and residential automation tools in the last 100 years and none of them are humanoid. I'd imagine a household robot butler would be more like R2D2 with lots and lots of arms.
It is when the world was made to interface with us. We can't use robots for everything if they aren't emulating us, because we would have to adapt everything for the non-humanlike robots.
Every single behavior? For sure not but otherwise we are the result of a very very long evolution and there is nothing else around us as smart and as adjustable.
The planing ahead thing through simulation for example seems to be a very good tool in neuronal network based architectures.
What if we put slop images into slop machines and got slop^2 back out
Maybe dumb question: One of the use cases is instrument reading of analog instruments. My brain immediately goes to "this should have some sensor sending data, and not be analog". Is having a robot dog read analog sensors really a better fit in some cases?
A parcel of land.
A few robot legs and arms, big battery, off-the-shelf GPU. Solar panels.
Prompt: "Take care of all this land within its limits and grow some veggies."
What if it turns out that "take care of this land" means the traditional way California was taken care of with regular small slow burns. After over 10k years of this type of management there are many important native species that won't even germinate without the presence of ash.
Or it could turn out to look like satayoma (Japanese peasant forests) or it could be more similar to the crop rotation that was traditionally practiced in many parts of Central Africa where roots were important.
In Russia before the Soviets forced "modern scientific agriculture" on peasants to modernize, they practiced things like contour farming (where they interplanted rows of crops against the contours of the land to slow water down) and maslins (where they intermixed multiple varieties of wheat and barleys in the same patch). Now contour farming are an active area of research for their ability to prevent topsoil loss and build soil health while maslins provide superior yield stability and use little to no pesticides.
That's not even getting into the over 40-120,000 varieties of rice we've documented. Most of which are hyper adapted to a very specific location—often even a single village.
My point is there is no one way to take care of a plot of land. It's all relative to a number of factors beyond just the abiotic characteristics of the land itself. Your goals and intentions matter and you will always find localized unique adaptations.
Yeah I'm not sure how that's currently working out. https://proofofcorn.com/
Are you saying that has failed? It isn't obvious to me from that page that anything in particular is going wrong. I don't think anyone is daft enough to claim that AI solves the "Iowa remains unplantable due to winter conditions" problem.
logs suggest it's been 'critically failing' and 'blocked for 68 days' on farmhand introduction, although the logs don't go back far enough (and cut off too early) to really tell what's going on. https://proofofcorn.com/log
Looking at the timeline they seem hopelessly behind. Its currently the planting window and they don't have land or a person to work it.
Ah, thank you. It's not the planting window yet where I am, north of Iowa, so I wasn't certain where they were.
That's kind of the opposite problem -- the agent doesn't have robot arms or legs or a parcel of land. It has to rely on people to get access to land and plant and harvest the corn, and those people are ignoring it.
Done! The whole planet is now veggies.
It turned the planet into ‘taterchips!
Google and Boston Dynamics (of Spot, Atlas fame) formed a partnership a while back and they’ve been working on building models together.
Hyundai now owns Boston Dynamics and is pushing to get the robots into their factories.
>Our safest robotics model yet Safety is integrated into every level of our embodied reasoning models. Gemini Robotics-ER 1.6 is our safest robotics model to date, demonstrating superior compliance with Gemini safety policies on adversarial spatial reasoning tasks compared to all previous generations.
The safety guidelines are interesting, they treat them as a goal that they are aspiring to achieve, which seems realistic. It’s not quite ready for prime time yet.
Pointing a camera at a pressure gauge and recording a graph is something that I would have found useful and have thought about writing. Does software like that exist that’s available to consumers?
I'm pretty sure claude will one shot this for you, including making you a home assistant dashboard item if you ask it.
Look into opencv.
frigate can be setup to do this I believe, but its overkill. Openclaw could do it, slightly less overkill.
This seems perfect to hook up to my 'LLMs can control robots over MCP' system. The idea is that LLMs are great at writing code, so let's lean in to that. I'll give it a try! I just got a bigger robot, we'll see how it does...
https://colinator.github.io/Ariel/post1.html
Really unfortunate that I forgot what YT video i saw about this just 2 weeks ago.
It was about Googles PaLM-E evolution and progress. It basically has two models one which controls the robot, the other is a llm and they are combined together in some attention layer.
Must have been https://www.youtube.com/watch?v=2mrGMMmrVNE from Welch Labs.
That video is pretty good. I'm basically betting that an earlier, abandoned approach described in the video, "Code as Policy", will beat everything else. It requires no training data, and generalizes instantly to all robots.
I’ve been thinking about AI robotics lately… if internally at labs they have a GPT-2, GPT-3 “equivalent” for robotics, you can’t really release that. If a robot unloading your dishwasher breaks one of your dishes once, this is a massive failure.
So there might be awesome progress behind the scenes, just not ready for the general public.
There's not enough internet-scale data for robotics. The gap is huge! So anyone that claims to have a GPT like model is not behing honest.
> If a robot unloading your dishwasher breaks one of your dishes once, this is a massive failure.
That's a bit exaggerated, no? Early roombas would get tangled in socks, drag pet poop all over the floor, break glass stuff and so on, and yet the market accepted that, evolved, and now we have plenty of cleaning robots from various companies, including cheap spying ones from china.
I actually think that there's a lot of value in being the first to deploy bots into homes, even if they aren't perfect. The amount of data you'd collect is invaluable, and by the looks of it, can't be synth generated in a lab.
I think the "safer" option is still the "bring them to factories first, offices next and homes last", but anyway I'm sure someone will jump straight to home deployments.
I ended up watching Bicentennial Man (1999) with Robin Williams over the weekend. If you haven't seen I thought it was a good and timely thing to watch and is kid friendly. Without giving away the plot, the scene where it was unloading the dishwasher...take my money!
From an economic standpoint the industry is anyway the most relevant by far. Its easier as the env is a lot more controlled, professionals configure and maintain the robots, they buy in bulk and have more money.
My concern with a household robot is not the dishwasher but the tv screen, the glas door, glas table, animals (fish/aquarium) etc. the robot might walk through, touch through or fall onto.
I have broken dishes loading and unloading the dishwasher. Am I a massive failure?
My non-AI dishwasher can't even always keep the water inside. Nothing is perfect.
If someone paid 100 grand for you to load and unload the the dishwasher, and the research to be able to do it costed hundreds of billions, decades of research, hundreds of thousands of researchers, and that was the ONLY thing you could do, yes, you WOULD be a massive failure.
> If a robot unloading your dishwasher breaks one of your dishes once, this is a massive failure.
Depending on what the rate of breaking dishes is, this would be a massive improvement on me, a human being, since I break a really important dish I needed to use like ~2x per month on average.
You really break a dish once every 2 weeks? That seems exceptionally clumsy.
Not here to shame you for it, for the record.
> That seems exceptionally clumsy
That's me ;_;
It's called "VLA" (vision-language-action) models: https://huggingface.co/models?pipeline_tag=robotics
VLA models essentially take a webcam screenshot + some text (think "put the red block in the right box") and output motor control instructions to achieve that.
Note: "Gemini Robotics-ER" is not a VLA, though Gemini does have a VLA model too: "Gemini Robotics".
A demo: https://www.youtube.com/watch?v=DeBLc2D6bvg
Meanwhile, gemini 3.1 pro (that was released two months ago) was completely unavailable to me this afternoon, neither with API nor Subscription.
Nothing was reported in Google status page, not even the CLI is responding, it’s just left there waiting for an answer that will never arrive even after 10 minutes.
Nice. I couldn't find the part that I'm most interested in though, latency. This beats their frontier vision model for some identification tasks -- for a robotics model, I'm interested in hz. Since this is an "Embodied Reasoning" model, I'm assuming it's fairly slow - it's designed to match with on-robot faster cycle models.
Anyway, cool.
In my quick image recognition testing on AI studio, it's performance seems similar to 3.1 pro, but is much much faster. It "thinks" but only for a few seconds.
Of course this is for counting animal legs while giving coordinates and reading analog clocks. Not coding or or solving puzzles. I imagine the image performance to model weight of this model is very high.
I thought the entire discourse on how important pointing is to be super interesting. I've been told, although I don't know if this is true, that dogs are the only animal that can understand human pointing. Fascinating to think this might be fundamental to world intelligence requirements. Well, it's required, but it's interesting to think that it might be a core structure required or that learning it might force some sort of neural architecture that's helpful.
And, I was disappointed to see that pointing was just giving x,y coords. I wanted to see robots pointing at stuff.
Would this approach destroy critical investments in physics- or modeling-based reasoning?
I'm all for the task reasoning and the multi-view recognition, based on relevant points. I'm very uncomfortable with the loose world "understanding".
The fault model I see is that e.g., this "visual understanding" will get things mostly right: enough to build and even deliver products. However, these are only probabilistic guarantees based on training sets, and those are unlikely to survive contact with a complex interactive world, particularly since robots are often repurposed as tasks change.
So it's a kind of moral-product-hazard: it delivers initial results but delays risk to later, so product developers will have incentives to build and leave users holding the bag. (Indeed: users are responsible for integration risks anyway.)
It hacks our assumptions: we think that you can take an MVP and productize it, but in this case, you'll never backfit the model to conform to the physics in a reliable way. I doubt there's any way to harness Gemini to depend on a physics model, so we'll end up with mostly-working sunk investments out in the market - slop robots so cheap that tight ones can't survive.
I wonder what new laws of physics AI models have discovered encoded opaquely in their weights, yet to be discovered by physicists
Is there a open source mini robot kit that allows me to play-around with agentic robots?
SO-ARM101 I guess? Or more likely the Lekivi variant.
I was also just in the market for a small experiment robot. I got the hiwonder armpi-fpv. Avoid it, the actuators are pretty bad - they're very 'grindy', the robot jitters like crazy when it moves. Any such problems with the lekivi?
Soon Open Source will fill the gap here as well
As the article notes regular Gemini and Gemma also have spatial reasoning capabilities, which I decided to test by seeing if Gemini could drive a little rover successfully, which it mostly did: https://martin.drashkov.com/2026/02/letting-gemini-drive-my-...
LLMs are really good at the sort of tasks that have been missing from robotics: understanding, reasoning, planning etc, so we'll likely see much more use of them in various robotics applications. I guess the main question right now is:
- who sends in the various fine-motor commands. The answer most labs/researchers have is "a smaller diffusion model", so the LLM acts as a planner, then a smaller faster diffusion model controls the actual motors. I suspect in many cases you can get away with the equivalent of a tool call - the LLM simply calls out a particular subroutine, like "go forward 1m" or "tilt camera right"
- what do you do about memory? All the models are either purely reactive or take a very small slice of history and use that as part of the input, so they all need some type of memory/state management system to actually allow them to work on a task for more than a little while. It's not clear to me whether this will be standardized and become part of models themselves, or everyone will just do their own thing.
For the fine-motor commands: or, the model can write the code to generate them on the fly. It seems to work, in my very limited experiment.
As for memory: my approach is to give the robot a python repl and, basically, a file system - the LLM can write modules, poke at the robot via interactive python, etc.
Basically, the LLM becomes a robot programmer, writing code in real-time.
Showing the murder dog reading a gauge using $$$ worth of model time is kinda not an amazing demo. We already know how to read gauges with machine vision. We also know how to order digital gauges out of industrial catalogs for under $50.
I think that where this gets interesting is when you can just drop these robotic systems into an environment that wasn't necessarily set up specifically to handle them. The $50 for your gauge isn't really the cost: it's engineering time to go through the whole environment and set it up so that the robotic system can deal with each of the specific tasks, each of which will require some bespoke setup.
Agree. I'm unclear what's the highlight of this post. Is the multimodality of the model (that can replace computer vision), is it the reasoning part, is it the overall wrapper that makes it very easy to develop on top?
It's the fact that is not task specific.
Completely agree, I get that this is a stepping stone for future, more reliable robots but I found the demonstration underwhelming.