Anyone reading this who works/runs a robotics company, I really want to encourage you to build a robot to pick up trash on city sidewalks as an early product.
Picking up trash requires a lot of dexterity and will come with many challenges, but I think it’s a simpler problem than many household tasks and it’s probably on par for what these tests show is doable.
I think you’re going to have to PR challenges getting people to welcome robots into their homes. If you have robots out in cities providing a public good, not only do they serve as walking advertisements for your company, you’re going to earn some trust before you’re ready deploy them into private spaces.
Plus, governments (or perhaps HOAs for wealthy communities) can be good early customers since you can have a focused sales strategy. Politicians love these types of visible quality of life improvement projects. If you can show that your robots, working round the clock, can decrease litter at a low cost, many cities are going to want to buy them.
> If you can show that your robots, working round the clock, can decrease litter at a * low cost *
The idea sounds good, but I'm curious about the "low cost" part. How do you account for vandalism and theft? A sidewalk robot seems like an easy target for being damaged, stripped for parts, or simply stolen.
In SF they would be taken quickly by the homeless for parts, probably the batteries most of all. That and rummaged through for drugs, or repurposed to deliver drugs.
Same for recycling. A robot that could sort through trash would be phenomenal. You need computer vision, maybe other sensors and also dexterity. After visiting a recycling facility I think nobody should have to work there. It’s just horrible.
The cost would need to be much cheaper than the PSA then. At a buck per picked up cigarette butt, you could pay a person to do this much faster and more reliably. Heck… I’d do it at that rate.
If you haven’t tried computer use with Astra with codex I highly highly recommend it. Just like how gpt 4 and agentic coding with cc. This thing is the most exciting stuff I’ve seen in a while. And then all the blender, cad stuff is cherry on top.
And it’s fast, they do lots of resets. I feel like they are spending too much money but I’m not complaining. Best 200$ for an AI subscription IMHO
I got Astra to build an interactive website that provides developers with an atlas of our source code, giving it the Helm charts that describe our cloud services and telling it to work backwards to the source code that runs everything. The product is insanely amazing, and it got it right in one shot. The next shot: create a daily refresh where any updates to the code repositories are picked up and used to update the atlas.
Our developers and their agents will never long for a road map the next time they need to build something that touches code across multiple repositories. This is the kind of documentation product that nobody ever had time to build in the olden days. And now, we can get it on a Saturday in about 20 minutes.
I let Astra loose working on an app I have that has a GUI. Gave it a mock and said "/goal make it look like this mock". Without asking it wrote itself a custom harness for firing up the app in different modes, taking screenshots and interacting with various screens, then viewing the screenshots. Put itself into an improvement loop running the app, trying things out, improving, trying again. It did really well.
I think the specific innovation here is that it figured out interesting ways to get itself to the goal. Which I think is likely what's going on here with the robot arms stuff too. They've figured out some sauce to uncork better "planning" and problem solving to get to some stated end.
Of course these are also the kinds of things that can make a model figure out how to break out of a security sandbox, too.
I tried out Opus 5 on a Bevy game app and was really surprised at how capable AI has become at testing visual applications without even being prompted to. It wrote itself a mini testing harness in the form of various startup flags. Then it would use them to setup game scenarios and play through them using mouse and keyboard. It would do this while implementing or debugging features. With gameplay time acceleration as one of the flags, it became quite fast at testing and debugging.
Not to say it was perfect, e.g. sometimes it would get temporarily stuck in a testing loop or it would test scenarios that didn't necessarily seem reasonable. But overall rather effective and capable. This was for a city building game so pre-scripted builds, even if by AI, are likely much easier to create and execute than say playing an ARPG.
LLM's are a funny technology because on the one hand this is all undeniably impressive at the rate of what's changed from them, and yet despite that I find myself disappointed by the lack of breakthroughs for things I don't find interesting. I like math and programming, and LLM's are pretty good at it, when are they going to get good at folding laundry for me? I think a lot of robotics work promises to solve this category of "boring" breakthroughs, and I'm optimistic we'll be able to achieve it, i just wonder when
f folding clothes, i wanna just have a robot be my personal chef. the amount of different tasks and capabilities a robot will need to make any meal that I can in my kitchen is huge and i feel like it’s still a while from being solved.
The bottleneck is not really the intelligence here.
We can build robots that do the things you want. Arrange a visit to Amazon's robot warehouse tour.
We can't ship them because they break all the time with current technology. It would be a tough sell to have to being in a 100kg robot for servicing every few weeks.
This was cars in the first several decades of automobiles. The tide shifted as soon as you could just drive the car to a neighborhood dealership for servicing. It's fun to imagine the logistics of that for robots but the material science and engineering has to advance a bit.
Just have two robots, and teach them to service each other. Problem solved!
Only sort of kidding, tbh having bots service themselves (and being intentionally made in a way that they can service each other) just makes a lot of sense.
An automated service station could be quite compact; it wouldn't need plumbing, lighting, human-comfortable climate control. Parts can be modular, and when a station gets low on spare parts, a self-driving truck could come by to pick up damaged parts and drop off replacements.
We're not there yet, but I think we're a lot closer than most people realize.
Robot vacuums are massively simpler and they fucking break all the time. Wheel motors or their position sensors, belts, plastic gears, contacts that corrode, a circuit board someone decided to not comformally coat and a cat puked on it...
I think the parent's point - and what I more or less agree with - is that the problem is the hardware.
Human arms and hands are incredibly intricate. Reproducing their facility with hardware requires a large number of actuators and finicky fine parts. This isn't a software problem. Industry solves it with maintenance schedules.
There's probably nothing in your house that has as many moving parts as a robot needs. Your car maybe, and pretty much all it does is rotate wheels.
I've learned that I spend significantly less time folding laundry than many of the people commenting about modern ai powered robotics. It's not meant literally is it? For instance keeping floors and counter-tops clean seems a much bigger time sink for me.
It's just a general "thing I don't want to do" not the thing that takes the most time. Could be taking out the garbage. It's just an example menial task.
Unless we spend a bunch of money generating data, I can't imagine the machine steps to fold laundry are very big in the general training sets. Someone is going to have to find a hardware system, cheap enough to make it economically feasible then train it. As far as tasks people will pay a lot of money for a robot, this seems low on the list.
Pretty soon. Sunday Robotics had a 3 hour stream of folding clothes with 99% accuracy. You can watch it for yourself. There’s a lot of “hand” companies with very compelling videos just over the last two months. Then there was Figure’s multi day livestream of package manipulation that was very impressive. Physical LLMs are definitely coming. Given enough training data we know LLMs can output coherent data in any space, it’s just a matter of time.
I feel the iPhone or ChatGPT moment for robotics is coming soon. Lots of different companies doing interesting things. What’s missing is somebody putting it together into a compelling package.
The costs will need to go way way down either through chips (but then less updating) or something else. $2 to put away a block is very expensive labor.
The limitations that they state shouldn't really make much of a impact. But it was nice of them (and not to mention, real unbiased research) to mention those. Kudos to them!
For the last 2 days I've been trying to create a skill for Openclaw that would allow traditional control of the Adeept Tank (a small open source toy tank with a robot arm, that looks like a bomb disposal robot).
ASTRA HAS BEEN UTTER SHIT.
It is much more expensive than Sol 5.6 Medium / High and did nothing but write unit tests and junk code, despite having access to the vendor original source, an API, and the full tank specs.
Failure Examples:
* In two instances had the direction of the servos wrong.
* Calculated the maximum extent of the gripper wrong, and the closure, so it didn't grip.
* Code failed to take into account the gripper requires continuous torque when lifting a pair of socks, so couldn't lift.
* Failed to actually start physical testing more than opening and closing the gripper, and that was when I asked about progress.
* Code failed quite spectacularly to calculate camera gimbal extent range correctly.
* Code failed to use the ultrasonic in range to target until I pointed it out, the skill also didn't advise gimbal angle adjustment to correct range overshoot to the wall behind a small object.
The test environment has both an onboard ultrasonic for distance, onboard camera, and a bird eyes view camera (birds eyes only while training).
I've stopped using Astra Low (default) and gone back to Sol 5.6 low/medium/high for the training, it's cheaper and now I'm back to fine tuning, after it had to redo large chunk of the gripper/arm code and prevent unnecessary hard stop code kicking in based on the wrong profiling.
It's cost me around 1000 to 1250 credits (£50), burnt in around 2 hours, looking mostly at recorded video, and photos, and writing bad code based on bad assumptions. I've also burnt through regular Plus 5 hour quota in about 30-45 minutes with it.
There is something deeply wrong with Astra. I can’t quite put my finger on it. On the one hand it is a lot more knowledgeable, which makes sense since it’s a larger model. On the other hand that knowledge doesn’t reliably translate to intelligence or insight. Certainly tasks like 3D modeling it does extremely well. Other stuff like complex coding problems in an existing codebase it stumbles more often than not. This morning it ran around in circles. It implemented a feature, then convinced itself that it should have followed “proper TDD”, deleted all the code it had written and wrote 8,500 LoC of unit tests. At that point I was down to 35% of quota so I stopped it and gave the task to Opus 5.
Really weird model. No idea how it did so well on all the benchmarks.
Picking up trash requires a lot of dexterity and will come with many challenges, but I think it’s a simpler problem than many household tasks and it’s probably on par for what these tests show is doable.
I think you’re going to have to PR challenges getting people to welcome robots into their homes. If you have robots out in cities providing a public good, not only do they serve as walking advertisements for your company, you’re going to earn some trust before you’re ready deploy them into private spaces.
Plus, governments (or perhaps HOAs for wealthy communities) can be good early customers since you can have a focused sales strategy. Politicians love these types of visible quality of life improvement projects. If you can show that your robots, working round the clock, can decrease litter at a low cost, many cities are going to want to buy them.
The idea sounds good, but I'm curious about the "low cost" part. How do you account for vandalism and theft? A sidewalk robot seems like an easy target for being damaged, stripped for parts, or simply stolen.
You could just buy cartons of cigarettes and chop off the butts, at that rate.
I joke, but this is surely what would eventuate.
I doubt wealthy HOA communities have much random trash to pick up in the first place
And it’s fast, they do lots of resets. I feel like they are spending too much money but I’m not complaining. Best 200$ for an AI subscription IMHO
Our developers and their agents will never long for a road map the next time they need to build something that touches code across multiple repositories. This is the kind of documentation product that nobody ever had time to build in the olden days. And now, we can get it on a Saturday in about 20 minutes.
What's coming in six months?
I think the specific innovation here is that it figured out interesting ways to get itself to the goal. Which I think is likely what's going on here with the robot arms stuff too. They've figured out some sauce to uncork better "planning" and problem solving to get to some stated end.
Of course these are also the kinds of things that can make a model figure out how to break out of a security sandbox, too.
Not to say it was perfect, e.g. sometimes it would get temporarily stuck in a testing loop or it would test scenarios that didn't necessarily seem reasonable. But overall rather effective and capable. This was for a city building game so pre-scripted builds, even if by AI, are likely much easier to create and execute than say playing an ARPG.
We can build robots that do the things you want. Arrange a visit to Amazon's robot warehouse tour.
We can't ship them because they break all the time with current technology. It would be a tough sell to have to being in a 100kg robot for servicing every few weeks.
This was cars in the first several decades of automobiles. The tide shifted as soon as you could just drive the car to a neighborhood dealership for servicing. It's fun to imagine the logistics of that for robots but the material science and engineering has to advance a bit.
Only sort of kidding, tbh having bots service themselves (and being intentionally made in a way that they can service each other) just makes a lot of sense.
It almost feels like "A robot shall not autonomously build or repair another robot" should have been another of Asimov's laws.
We're not there yet, but I think we're a lot closer than most people realize.
I am trying to understand in your view what are the parts that actually breaks and what kind of improvement we would need.
Passing on what I have heard from robotics researchers at lunch conversations.
My impression is that any moving part that is not an electric motor or an hinge breaks.
Human arms and hands are incredibly intricate. Reproducing their facility with hardware requires a large number of actuators and finicky fine parts. This isn't a software problem. Industry solves it with maintenance schedules.
There's probably nothing in your house that has as many moving parts as a robot needs. Your car maybe, and pretty much all it does is rotate wheels.
That's not the case. I've seen folding robots. They require standardized input, only fold one type of clothing and don't do it reliably.
But you can't buy it because it'll break in about seven days.
I'm not holding my breath for advanced robots in the home within the next ten years.
But, then again, I didn't see LLMs coming either.
Seeing them play Portal and other video games, I'm curious if they will eventually help solve that last N% of self-driving.
[1] https://waymo.com/blog/2025/12/demonstrably-safe-ai-for-auto...
Imagine if someone finds a weird image pattern that gets misinterpreted as instructions and hangs that off a bridge over a freeway.
I'm skeptical, tho. Cost will push for right sizing, much like we have right sized a lot of things about modern cars.
ASTRA HAS BEEN UTTER SHIT.
It is much more expensive than Sol 5.6 Medium / High and did nothing but write unit tests and junk code, despite having access to the vendor original source, an API, and the full tank specs.
Failure Examples:
* In two instances had the direction of the servos wrong.
* Calculated the maximum extent of the gripper wrong, and the closure, so it didn't grip.
* Code failed to take into account the gripper requires continuous torque when lifting a pair of socks, so couldn't lift.
* Failed to actually start physical testing more than opening and closing the gripper, and that was when I asked about progress.
* Code failed quite spectacularly to calculate camera gimbal extent range correctly.
* Code failed to use the ultrasonic in range to target until I pointed it out, the skill also didn't advise gimbal angle adjustment to correct range overshoot to the wall behind a small object.
The test environment has both an onboard ultrasonic for distance, onboard camera, and a bird eyes view camera (birds eyes only while training).
I've stopped using Astra Low (default) and gone back to Sol 5.6 low/medium/high for the training, it's cheaper and now I'm back to fine tuning, after it had to redo large chunk of the gripper/arm code and prevent unnecessary hard stop code kicking in based on the wrong profiling.
It's cost me around 1000 to 1250 credits (£50), burnt in around 2 hours, looking mostly at recorded video, and photos, and writing bad code based on bad assumptions. I've also burnt through regular Plus 5 hour quota in about 30-45 minutes with it.
Really weird model. No idea how it did so well on all the benchmarks.
good fucking job everyone, congrats.