[iandanforth]

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).

[duchenne]

I am working in Mistral robotics team. I confirm this is map-less. The only inputs are the text prompt and the front camera rgb image.

[vga1]

Please oh please try to make Kärcher adopt your stuff. Even their latest premium cleaning bots are hopeless when they don't know where they are, even when I tell them where they are.

Or, I don't know, make your own vacuums.

[dylan604]

I have no hope for this working. I recently bought a dumb meatbag operated vacuum after tiring of the robot's issues. The constant getting stuck in places drove me crazy. After getting a new vac, I used it in a place where the robot had just "cleaned" that morning. According to the new collector's contents, the robot sucked at its job of sucking.

[vga1]

The sad part is that my previous robot (iRobot model from 3 years ago) was better at this.

[dylan604]

Mine worked pretty well as well and slightly improved after an update, but has gotten dumber over the years. Not only does it get lost, but it physically gets stuck under things that it did not used to get stuck under.

[moffkalast]

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?

[mike_hearn]

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".

[moffkalast]

Yeah the benefits of a model that can understand text well enough is certainly that one could in theory just describe the layout. Just let it know what connects to what, what's on each floor, etc. Might work if it can figure out where it is based on contextual info, but probably not super well yet I imagine. VLAs are known to have very little text context, like sentences worth at most, maybe this one's different.

[jgrizou]

Do you think a similar model architecture could include some sort of memory, so a follow-up prompt of "Go back to the water foutain" would somehow work?

A in-model memory approach is probably still deep research but maybe a Rag-like pipeline could work in some instances

[rao-v]

Could you please open source this or a 4B version? I’ve been messing around w hooking up vllms to cheap robots and skipping the whole ROS stack and this would be an absolute delight to play with

[iandanforth]

Nice! Thanks for confirming. If the last instruction is 'go back to where you started' does it work?

[duchenne]

Yes, it would work.

[digitaltrees]

Are you interested in working with partners that are collecting training data?

[ainch]

Wouldn't modern SLAM or VSLAM address that problem?

[pj_mukh]

Would be extremely interesting to build an "Exploration" node of sorts. Solve a sort of Semantic SLAM problem as you go.

So if you can give it an instruction to "Find the elevator on this floor", could it walk around and build a map as it goes so it starts doing what a human would do to find the elevator.

As of now, the way these navigation models are setup, it assumes the instruction writer was intimately aware of useful visual navigational landmarks to give, which is not realistic for most use cases.

[whiplash451]

https://dspace.mit.edu/bitstream/handle/1721.1/73150/Teller_...

ICRA, 2010

[mostlyk]

Can you explain how it is much easier if you have a pre-captured map given what they are doing without using any sensors, all you have is perhaps these recent feed forward tokens not actual Geometry.