It's implied, and I'm hoping it's true, that this is a map-less navigation. Which is impressive. This kind of task is much easier if you have a pre-captured map of the environment, but if they are doing this without a map it's great. Historically you were always faced with "The Kidnapped Robot" problem where robots that didn't know where they were couldn't navigate even a little bit. Here the robot appears to be able to follow directions as long as they are interpretable from its current vision (or via dead reckoning).
So, it has to rely on exact situational step by step commands? I'm wondering how one could conceivably deploy this in a useful way. Usually you'd need to mark areas on the map and then the robot knows where to go, A* is trivial around obstacles once you have that and a lidar. And lidars are an order of magnitude cheaper than something that can run an 8B VLA.
One could maybe autogenerate these text planning commands, but it would require a map and the robot's current location, so it doesn't really solve that, unless it can find a specific thing completely on its own. How much of a planning horizon does it have?
You probably don't need a geometric map. Just have someone wander around with a mobile app and feed the video into a more powerful model once, asking it to produce descriptions of the different areas of the office or building and how they connect. Now you have a "text adventure game" map you can use it with a small LLM to produce instructions for the robot to follow, assuming it knows where it currently is.
The advantage over traditional approaches is presumably flexibility. LIDAR isn't going to solve an instruction like "find the man with the pink shirt".
This looks to not be an openly available model, but I think if it were, availability of an easy single-camera navigation setup could allow for a lot of cool hobbyist projects.
I wonder how Mistral will prioritize its robotic development against its LLM development. We have either players that prioritize both (Google, AMI), or players that prioritize coding and agentic (OpenAI, Anthropic, ...).
If you're wondering what prevents or mitigates AI hallucinations on the AI layer from replicating or acting out on the physical layer look up QNX. They manage the deterministic reasonin gof robotics. You know them better as Blackberry.
"Go to the next room" and there is two doors, what do you do ?", "turn at the water dispenser" and there is a sink, that sort of things I assume is the biggest thing they're facing (beside the last 1% that's worth another 99%, as usual).
On their page where the result graph is, go to navigation error, that's the one that matters for your question, and you see their model is great at not navigating "wrong", so their failure rate was that it couldn't figure it out.
Probably it achieved outside-from-outside in discrete void. Teleportation wasn’t an expected outcome for this experiment, but on the other hand the instructions didn’t forbid that kind of move.
This is very cool. Congratulations to the Mistral team. Map less navigation in the outside world has been around for quite a while. But map less navigation inside the buildings is relatively new. Some stanford researchers trained a vision model (PIGEON) which could tell the geo-location from any image. It was not released publicly due to privacy nightmarish (stalking!) possibilities but I am assuming similar type of tech has gone behind this robot. if someone knows more, feel free to correct.
Ok, this is really cool. The fact that the robot can use pointing to decide where to go is a great design decision, and robotics really is the next frontier. Definitely cheering on Mistral here!
For a claim such as state of the art, or claims such as "great at any task" needs something of more substance. I've seen maze-solving robot competitions which can zoom around in seconds. The sped up video in the first part, and the "obstacle avoidance" are too slow for me to believe this is state of the art.
While impressive at 8B, what would the expectation be in real life, that it's run remotely or autonomously with a strapped on GPU and battery?
I've used that example as a contrast of what I've seen before. If you can point me at comparable efforts, in the same category as what Mistral is doing, I'd be interested in having a comparative look.
All I can think of are robot dogs, Tesla bots, and whatever flavor of the month Japanese robots show up at trade shows.
It's potentially a great strategy. They can't keep up with Antropic and OpenAI in pure horsepower, but there's just tons of applications for which you don't need that much power and it's better to optimize for speed and energy.
Producing specific niche models for 100 year old industries that have mountains of data and warehouses full of folders will be the european take on AI.
It may come late but it‘ll be safe and reliable. It also requires a lot of OCR.
The Niche model story is still fairly week. Evidence points to general models being equally capable to niche models at a more attractive capex (risk is spread across multiple verticals rather than concentrated in a single model capability)
The General Models' business-model is also looking more weak every iteration.
Costs of simple tasks grow extensively: OCR with "Mistral OCR" at $4 per 1000 pages vs OCR with Opus 4.8 at sometimes¹ $1 per "page".
Or just the immense costs when burning tokens in an unoptimized agentic coding environment costing tens of dollars for a few simple classes or functions versus a highly optimized "autocomplete" model costing under $10 for thousands of such classes and functions.
Or the, over ten dollars worth of tokens when some "agent" using a general model, tries to perform the task I gave it to "read the event on example.com/event/1337 and put it in my calendar", include commute time as well"
The "general models" currently only become smarter by growing bigger and having larger context windows - by becoming exponentially more expensive to train and to run and to interact with. Whereas "Niche" models can do the things that "normal code" cannot do, and improve by tuning and tweaking only that. Their goal is then to fill in gaps that traditionally are hard or impossible with normal software. Wheras the goal of a general model (with agentic reasoning)is to replace that entire "normal software".
One example: I am not interested in "chatting with my calendar". I'm interested in a calendar because it is a well known view (UI) of my planning and tasks, but I see a lot of opportunities where AI can improve my working with this calendar. I may be interested in a smarter screen when I hit "+ Add event"; one that has knowledge of my previous events and patterns (some RAG vector db maybe). One that maybe has access to content I just copied, or read (though: privacy?) or can open my camera to let me shoot a pic of something that has the event info on it.
In such a set-up, Niche LLMs perform dedicated tasks: determine patterns (he always books a Yoga class on wednesday or thursday, two days in advance, so lets suggest a yoga class), determine existing content (event is planned 100Km from his home, so lets suggest the commute based on previous commutes like this). Or an OCR model. Or an autocomplete model. Relatively simple, niche models, called from within software to aid me when "calendaring". Not replace the entire calendar with some chat.
¹Edit: This was a rather unscientific research of mine, where I compared some models to read from photographs, compared purely on costs and timing. "Opus" or other generic LLMS with image input capabilities commonly did better on "performance" esp with difficult input such as a picture of a poster of some rock event.
The cost is getting worse and worse for large general models, they're already way past that point in economics. Also, mMistral specialize in "on site" models, not remote. In terms of capex, renting factory/warehouse/whatever robots versus buying them and depreciate has already been played out, companies didn't want to replace human employees with robots employees.
It seems like a stronger story for robotics, since smaller models can always react to the environment faster than large models at a given hardware budget. Also because robots that keep their models local for latency or reliability aren't going to be carrying many kilowatts of inference capacity.
There are many, many factories that still don't have internet access on the floor, and commercial inference generally has response latencies measured in seconds. I struggle to imagine a factory spending hundreds of thousands for the local compute to run a large model either, given how cheap they are about expenses.
I'm also skeptical that you can cleanly differentiate between "safety critical actions" and "actions", though this is less of a practical concern given how laissez-faire some manufacturers are. For context, I work on safety critical robotics (in automotive).
We are making a niche model that we're now expanding. You'd be surprised how the general models suck for anything vision-related.
And even if you use all the tricks in the book to make them work for you, the cost can easily be 1000 _times_ more than the specialized model. Ditto for speed.
This is especially important for things like robotics or navigation.
It follows directly from the bitter lesson - a frontier model can be relatively cheaply distilled into anything you need to run quickly (and a frontier model like Mythos will help you distill it quickly), decidedly not true the other way around.
Funny how nearly all model improvements this year are demonstrated on the subset of use cases where brute force / reinforcement learning is most effective:
Robotics (using physics sims)
Cybersecurity (red team / blue team)
Math (using automated proof checkers)
Programming (using compilers)
For the record I think robotics is a totally logical place to use this training approach and this is very impressive. But if we zoom out and think about LLMs in general I’m not sure this inspires confidence in AGI arriving any time soon. I would also propose that this is a form of overfitting / training-test contamination.
Take cybersecurity for example. Through brute force techniques you will gradually memorize all of the possible exploits. So when fable breaks into a DoD network everyone is shocked but in reality it basically memorized all possible exploits including some zero day.
I’d be much more interested to see if fables performance is preserved as new exploits arise (NOT zero day - negative day meaning exploits that don’t exist yet). Would fable still find them? Or would they need to retrain it on the new software stack continuously in order to identify the zero days.
This is an important distinction that I have not seen made before.
It is already here. Not humanoid (yet, but it's in the works) but tracked robots with bolted on machine guns have both held and captured positions in UA.
It’s unclear to me what their desired outcome for a blog post like this. If you’ve ever worked in a robotics setting, 80% implies that 20% of your autonomous actions are incorrect. Imagine if this were the case for autonomous driving where your car misbehaves 1 in every 5 actions it takes.
Posts like this just reminds me of the end to end demos AV companies built in the early days using a single camera - only to realize that it’s harder than it looks years later into development.
The ICP question was more around the model itself. Are they looking to license it to robotics companies? Do they imagine that devs at robotics companies would be willing to deploy these models as a black box?
One could maybe autogenerate these text planning commands, but it would require a map and the robot's current location, so it doesn't really solve that, unless it can find a specific thing completely on its own. How much of a planning horizon does it have?
The advantage over traditional approaches is presumably flexibility. LIDAR isn't going to solve an instruction like "find the man with the pink shirt".
I would like to know what it did the other 23.4% of the time!
On their page where the result graph is, go to navigation error, that's the one that matters for your question, and you see their model is great at not navigating "wrong", so their failure rate was that it couldn't figure it out.
[/joke]
here's the link to the PIGEON paper - https://lukashaas.github.io/PIGEON-CVPR24/
While impressive at 8B, what would the expectation be in real life, that it's run remotely or autonomously with a strapped on GPU and battery?
All I can think of are robot dogs, Tesla bots, and whatever flavor of the month Japanese robots show up at trade shows.
It may come late but it‘ll be safe and reliable. It also requires a lot of OCR.
Costs of simple tasks grow extensively: OCR with "Mistral OCR" at $4 per 1000 pages vs OCR with Opus 4.8 at sometimes¹ $1 per "page".
Or just the immense costs when burning tokens in an unoptimized agentic coding environment costing tens of dollars for a few simple classes or functions versus a highly optimized "autocomplete" model costing under $10 for thousands of such classes and functions.
Or the, over ten dollars worth of tokens when some "agent" using a general model, tries to perform the task I gave it to "read the event on example.com/event/1337 and put it in my calendar", include commute time as well"
The "general models" currently only become smarter by growing bigger and having larger context windows - by becoming exponentially more expensive to train and to run and to interact with. Whereas "Niche" models can do the things that "normal code" cannot do, and improve by tuning and tweaking only that. Their goal is then to fill in gaps that traditionally are hard or impossible with normal software. Wheras the goal of a general model (with agentic reasoning)is to replace that entire "normal software".
One example: I am not interested in "chatting with my calendar". I'm interested in a calendar because it is a well known view (UI) of my planning and tasks, but I see a lot of opportunities where AI can improve my working with this calendar. I may be interested in a smarter screen when I hit "+ Add event"; one that has knowledge of my previous events and patterns (some RAG vector db maybe). One that maybe has access to content I just copied, or read (though: privacy?) or can open my camera to let me shoot a pic of something that has the event info on it. In such a set-up, Niche LLMs perform dedicated tasks: determine patterns (he always books a Yoga class on wednesday or thursday, two days in advance, so lets suggest a yoga class), determine existing content (event is planned 100Km from his home, so lets suggest the commute based on previous commutes like this). Or an OCR model. Or an autocomplete model. Relatively simple, niche models, called from within software to aid me when "calendaring". Not replace the entire calendar with some chat.
¹Edit: This was a rather unscientific research of mine, where I compared some models to read from photographs, compared purely on costs and timing. "Opus" or other generic LLMS with image input capabilities commonly did better on "performance" esp with difficult input such as a picture of a poster of some rock event.
Unless you are in military robotics or automotive of course :)
I'm also skeptical that you can cleanly differentiate between "safety critical actions" and "actions", though this is less of a practical concern given how laissez-faire some manufacturers are. For context, I work on safety critical robotics (in automotive).
And even if you use all the tricks in the book to make them work for you, the cost can easily be 1000 _times_ more than the specialized model. Ditto for speed.
This is especially important for things like robotics or navigation.
Robotics (using physics sims)
Cybersecurity (red team / blue team)
Math (using automated proof checkers)
Programming (using compilers)
For the record I think robotics is a totally logical place to use this training approach and this is very impressive. But if we zoom out and think about LLMs in general I’m not sure this inspires confidence in AGI arriving any time soon. I would also propose that this is a form of overfitting / training-test contamination.
Take cybersecurity for example. Through brute force techniques you will gradually memorize all of the possible exploits. So when fable breaks into a DoD network everyone is shocked but in reality it basically memorized all possible exploits including some zero day.
I’d be much more interested to see if fables performance is preserved as new exploits arise (NOT zero day - negative day meaning exploits that don’t exist yet). Would fable still find them? Or would they need to retrain it on the new software stack continuously in order to identify the zero days.
This is an important distinction that I have not seen made before.
This analysis by Toby Ord demonstrates why it’s a problem if frontier improvements are coming from reinforcement learning (brute force methods) from a purely computational perspective: https://www.tobyord.com/writing/inefficiency-of-reinforcemen...
But I'm scared for when those home helpers get drafted to fight in wars, either for or against me...
I imagine the EU would block any attempted takeover of Mistral given recent Anthropic and US govt actions.
SOTA 80% means a practically useless robot. What are they really imagining their ICP to be here?
Posts like this just reminds me of the end to end demos AV companies built in the early days using a single camera - only to realize that it’s harder than it looks years later into development.