Robot dogs now read gauges and thermometers using Google Gemini

Google's AI enables robots to read gauges while inspecting industrial facilities.

See full article...

 

[UserIDAlreadyInUse]

At the very least, the latest model may nudge us one step closer to a future where a General Atomics International Mark 4 robot can scan the room and correctly exclaim, “There’s no fudge here!”

Before or after it opens fire following a failed Hack attempt?

 

[TehRoot]

Pretty cool. Can see a lot of utility for keeping humans out of extremely dangerous spaces whenever possible except in shutdown scenarios (maintenance).

Have concerns over the model accuracy with reading certain things in real time and making interpretations, but there have been many cases where humans have ignored or misinterpreted similar information and it killed them or others with them, at least in scenarios like that the only thing at risk immediately is the machine.

Probably a net good here.

 

[DarthSlack]

Those improvements come courtesy of Google DeepMind’s newest robotic AI model that aims to enhance robotic capabilities for ‘embodied reasoning’ when interacting with physical environments.

Or, perhaps (with apologies to Douglas Adams)......

"Here I am, brain the size of a planet, and they ask me to read a bunch of dials and gauges. Call that job satisfaction? 'Cos I don't"

 

[sandragen]

But how reliably?

 

[pug fugly]

What could possibly go wrong?

 

[GorillaInAHumanSuit]

But how reliably?

86 percent of the time, it hallucinates less than 10 percent of the time

 

[TheOldChevy]

Not sure I fully understand the use case :

The new model can also more accurately perceive the risk of injury to humans in different scenarios, such as a young child sticking something into an electrical socket.

Is this nice dog supposed to autonomously monitor young childs?

I don't know if I would prefer a Doberman for that same purpose.

 

[ubercurmudgeon]

Is it housebroken, or will it be leaving batteries all over the place?

 

[Me Myself And I1]

Fun fact. Remote heating - a typical feature of many European cities - requires meters to be installed on actual individual radiators in every apartment.

Those meters have been reliably 'read' remotely for over 20 years. No need for physical reading and entry.

And I'm guessing that 'dangerous environments' deployed a form of remote reading/monitoring long time ago.

This is a PR stunt, a solution in search of a problem.

 

[randomcat]

maintenance managers are about to discover that every blower ever installed in every manufacturing facility in the world badly needs its filter replaced, assuming the robots have some kind of end effector to brush the dust off gauges that have never been checked before

 

[FinWiz]

But how reliably?

From the article:

The agentic vision capability reportedly boosts robotic performance on instrument reading tasks from 23 percent in the older Gemini Robotics-ER 1.5 model to 98 percent in the new Gemini Robotics-ER 1.6 model. For comparison, Gemini 3.0 Flash delivered just 67 percent accuracy. The baseline Gemini Robotics-ER 1.6 model can still achieve 86 percent accuracy in reading instruments even without agentic vision.

98% is very impressive considering where we were at only a few years ago. One mistake in 49 tasks. If running a second model could produce an independent test (which is far from certain, mind you), running two checks would give you a 0.04% error rate, or 1 in 2,500.

A Canadian pharmacist friend told me that the average error rate for patient-received prescriptions is approximately 17k in 91 million, or 1 in 5168. Two of those errors resulted in death; 87 in moderate to severe harm. While there's much more to dispensing than simply counting pills, if similar tasks could be meaningfully double-checked by models, that'd be an enormous boost to the health system.

 

[mangoslice]

Google also describes Gemini Robotics-ER 1.6 as its “safest robotics model yet,” with a “substantially improved capacity to adhere to physical safety constraints.”

Not the main point of the article and it’s been discussed before, but these types of “best EVER” statements (which Apple always includes in its product announcements) still drive me bonkers.

 

[Hypatia]

[…]That implies the newer model has less of a “hallucination” problem than the older one, even if the latest model is still far from achieving human-level comprehension of its surroundings.

Does it? Without quite a bit more information on how the test was constructed and how the new training/programming has been done, I have little confidence in the implications of this.

After all, it appears that the marketing hype of all things “AI” has been thick and deep and is approaching propaganda levels of nonsense: https://www.flyingpenguin.com/freebsd-cve-2026-4747-log-suggests-mythos-is-a-marketing-trick/

 

[Fatesrider]

That implies the newer model has less of a “hallucination” problem than the older one, even if the latest model is still far from achieving human-level comprehension of its surroundings.

I'd go with "identification", rather than comprehension.

Comprehension is grasping with the intellect - aka knowing. Lacking intellect, it can't know. It simply aligns previous input with a preponderance of observed data, and then synthesizes a statement that relates the relationship is as close to that alignment as it concluded based on that preponderance of observed data.

Sophisticated, yes. Intellect - far from it.

 

[randomcat]

While there's much more to dispensing than simply counting pills, if similar tasks could be meaningfully double-checked by models, that'd be an enormous boost to the health system.

That would be true if the system worked perfectly and only corrected real errors. In practice these systems have lots and lots of false positives that then require extra resources to verify, so you have to constantly manually check the thing that's supposed to be automaticaly checking things for you.

 

[DistinctivelyCanuck]

and how many industrial jobs just vanished ?

(it was a coin toss between that, and "is there a T-800 model coming?")

 

[SomewhereAroundBarstow]

Ruh ro.

 

[randomcat]

I'd go with "identification", rather than comprehension.

Comprehension is grasping with the intellect - aka knowing. Lacking intellect, it can't know. It simply aligns previous input with a preponderance of observed data, and then synthesizes a statement that relates the relationship is as close to that alignment as it concluded based on that preponderance of observed data.

Sophisticated, yes. Intellect - far from it.

This is an important distinction. Despite impressive performance--sometimes, anyway--what is currently being sold as "AI" is almost never acting logically on an internal model of the world, as most people probably assume. It's a vector translation of the question made to look like a plausible answer. Sometimes that's all you need, but overselling it is going to get a lot of people killed.

 

[Robin-3]

Robots such as Boston Dynamics’ four-legged Spot can now accurately read analog thermometers and pressure gauges while roaming around factories and warehouses. Those improvements come courtesy of Google DeepMind’s newest robotic AI model that aims to enhance robotic capabilities for ‘embodied reasoning’ when interacting with physical environments.

<...>

The agentic vision capability reportedly boosts robotic performance on instrument reading tasks from 23 percent in the older Gemini Robotics-ER 1.5 model to 98 percent in the new Gemini Robotics-ER 1.6 model. For comparison, Gemini 3.0 Flash delivered just 67 percent accuracy.

98% accuracy is pretty nifty.

But, like everything else AI-based, is it also absolutely confident when it's wrong? Do its errors skew towards what's "expected" or "predictable" (see: all the analogies to super-advanced predictive text) when that disagrees with what it's actually reading? Either or both of these seem likely, and also problematic.

Also, on a more existential level: there are SO many more fields where I'd be happier to think "what an incredible new development!" than ai-enabled robot "dogs" (especially those already being adapted for police/military/security work).

 

[GFKBill]

98% is very impressive considering where we were at only a few years ago. One mistake in 49 tasks. If running a second model could produce an independent test (which is far from certain, mind you), running two checks would give you a 0.04% error rate, or 1 in 2,500.

A Canadian pharmacist friend told me that the average error rate for patient-received prescriptions is approximately 17k in 91 million, or 1 in 5168. Two of those errors resulted in death; 87 in moderate to severe harm. While there's much more to dispensing than simply counting pills, if similar tasks could be meaningfully double-checked by models, that'd be an enormous boost to the health system.

Did you really just compare a system that makes a mistake 1 in 50 (98% is not 1 in 49 fyi) to humans achieving 1 in 5168?

Hypothesising their model somehow jumping two orders of magnitude is pretty questionable. Like for like, they went from 23% (yikes!) to 67% 86% in a year, and adding agentic on top got them to 98%.

Getting to 99.98% on those prescriptions is a huge gap. Diminishing returns etc.

Edit: Stat correction

 

[GFKBill]

and how many industrial jobs just vanished ?

(it was a coin toss between that, and "is there a T-800 model coming?")

None.

Nobody is buying an expensive robo-dog that is that innacurate to monitor the systems they already have a perfectly reliable way of monitoring now.

 

[GFKBill]

Google also describes Gemini Robotics-ER 1.6 as its “safest robotics model yet,” with a “substantially improved capacity to adhere to physical safety constraints.”

Not the main point of the article and it’s been discussed before, but these types of “best EVER” statements (which Apple always includes in its product announcements) still drive me bonkers.

"Our new scaffolds are 30% less likely to randomly disintegrate!"

Call me when you're able to advertise with words like "reliable", "certified", "guaranteed" etc

 

[dtich]

"Temperature at Critical, must vent reactor core--"

"SPOT!! NO! That's not the reactor temp, that's the water flow rate! DO NOT VENT!"

"CRITICAL TEMPERATURE DETECTED - VENTING REACTOR CHAMBER - WARNING WARNING!"

"SPOT!! STOP!!!"

 

[TheBaconson]

I wonder if its been trained to know about parallax error.

 

[jaqrah]

It needs a pincher arm to drag incapacitated humans away from dangerous situations.

That is an application I could get behind.

 

[Maxipad]

Fun fact. Remote heating - a typical feature of many European cities - requires meters to be installed on actual individual radiators in every apartment.

Those meters have been reliably 'read' remotely for over 20 years. No need for physical reading and entry.

And I'm guessing that 'dangerous environments' deployed a form of remote reading/monitoring long time ago.

This is a PR stunt, a solution in search of a problem.

Yah. Normally all of these things are read using telemetry in any large or even small plant. I'd never trust this thing to do the job more reliably or better.

 

[fenwick67]

why not just spend $100 on a connected appliance for your industrial thermometer instead of $75k for a robot dog that takes 80 minutes to charge and 10 minutes to give you a reading

 

[NameRedacted]

Those meters have been reliably 'read' remotely for over 20 years. No need for physical reading and entry.

And I'm guessing that 'dangerous environments' deployed a form of remote reading/monitoring long time ago.

Kind of. Those are reading in a very specific range and in climate-controlled situations. There is plenty of remote instruments used in industrial facilities (and for a lot longer than 20 years!) but there’s a core issue that they may break and they may read incorrectly and you may not know. I work in risk assessment and we typically see a failure rate of 1-10% per year. In many cases the only way you can know it has failed is by going out and confirming with a gauge.

That is the “supposed” value of operator rounds: confirming that what is being read into the computer is reality. But there is a second reason we do them: people are good at noticing when something sounds/looks/smells funny, and they can’t do that without being out there. Most major incidents are stopped by someone going “huh, that doesn’t look/sound/smell right”. The robots aren’t doing that, so the benefit is lost.

The real reason this is just a marketing gimmick is that the robot is using a camera— A camera that could just as easily be shown on a monitor and have a real human look at it. The “AI” is not actually doing anything useful

 

[DistinctivelyCanuck]

None.

Nobody is buying an expensive robo-dog that is that innacurate to monitor the systems they already have a perfectly reliable way of monitoring now.

Do you seriously think there aren't MBA a**holes' who aren't going to look at this thing and consider "don't have to staff a position, don't have to pay it benefits, don't have to worry about sick days etc. etc. " and instantly justify a business case for it?

I've worked with MBA A**holes who would sell their mothers to find another quarter point of profit: so yeah, I can see these things being looked at as an easy way to chop headcount.

 

[TheBaconson]

Kind of. Those are reading in a very specific range and in climate-controlled situations. There is plenty of remote instruments used in industrial facilities (and for a lot longer than 20 years!) but there’s a core issue that they may break and they may read incorrectly and you may not know. I work in risk assessment and we typically see a failure rate of 1-10% per year. In many cases the only way you can know it has failed is by going out and confirming with a gauge.

That is the “supposed” value of operator rounds: confirming that what is being read into the computer is reality. But there is a second reason we do them: people are good at noticing when something sounds/looks/smells funny, and they can’t do that without being out there. Most major incidents are stopped by someone going “huh, that doesn’t look/sound/smell right”. The robots aren’t doing that, so the benefit is lost.

The real reason this is just a marketing gimmick is that the robot is using a camera— A camera that could just as easily be shown on a monitor and have a real human look at it. The “AI” is not actually doing anything useful

I know what you mean about the human side, back in my younger days in industrial maintenance, walking out to one of the factory floors smelling the air and knowing I’ll be replacing a motor soon, I can smell the cooking windings.

Edit; which then meant I could go track it down and schedule in the repair with the production team, as to cause minimal disruption.

 

[JMTronicHobbyist]

Fun fact. Remote heating - a typical feature of many European cities - requires meters to be installed on actual individual radiators in every apartment.

Those meters have been reliably 'read' remotely for over 20 years. No need for physical reading and entry.

And I'm guessing that 'dangerous environments' deployed a form of remote reading/monitoring long time ago.

This is a PR stunt, a solution in search of a problem.

If only they would carry a pencil and clip board to mark the readings on an invoice. At least people could maybe have some fun watching them, like some people get excited to see different trains.

 

[GFKBill]

Do you seriously think there aren't MBA a**holes' who aren't going to look at this thing and consider "don't have to staff a position, don't have to pay it benefits, don't have to worry about sick days etc. etc. " and instantly justify a business case for it?

I've worked with MBA A**holes who would sell their mothers to find another quarter point of profit: so yeah, I can see these things being looked at as an easy way to chop headcount.

I hear you, but we're talking industrial process control type environments. The maintenance/monitoring worker is pocket change compared to having production go down. Not even an MBA is dumb enough to risk millions in lost output by sticking a 98% accurate robodog in there instead.

And as I mentioned earlier, these companies are likely already highly automated and have electronic remote monitoring. The overlap between "can afford a robodog (and someone to watch and maintain robodog)" and "have manual gauges and shizz we need someone to walk around looking at" I would wager is pretty small.

 

[ngoncalves]

Fun fact. Remote heating - a typical feature of many European cities - requires meters to be installed on actual individual radiators in every apartment.

Those meters have been reliably 'read' remotely for over 20 years. No need for physical reading and entry.

And I'm guessing that 'dangerous environments' deployed a form of remote reading/monitoring long time ago.

This is a PR stunt, a solution in search of a problem.

A PR stunt sure, but the article always mentions factory environments. Here in Belgium I know AB inBev (beer producer) uses this robot exactly for this task (inspecting the facility to improve planned maintenance jobs) at its factory in Leuven. Is this more efficient than installing a network of meters and sensors all over the factory floor ? I would say it depends on the type of data you need.

Anyway, I don't see this becoming a reality in domestic environments because homes are highly caotic environments, when compared with a factory. And a robot takes up space and requires maintenance, yet another home appliance you will need to think about.

 

[clb2c4e]

Fun fact. Remote heating - a typical feature of many European cities - requires meters to be installed on actual individual radiators in every apartment.

Those meters have been reliably 'read' remotely for over 20 years. No need for physical reading and entry.

And I'm guessing that 'dangerous environments' deployed a form of remote reading/monitoring long time ago.

This is a PR stunt, a solution in search of a problem.

And at an error every 50 checks, statistically, where there may be more than 50 meters to check for it to be worthwhile, it not only complicates but introduces significant error.

 

[randomname19937]

But how reliably?

How reliably do humans? I've once received a horrendous gas bill because gas company worker misread 1300 as 1800 (that's in cubic meters)...

 

[timaua]

Query: Is that a stick?

Answer: Oh!, yeah!, yeah! Yes, stick!! It’s a stick, it’s a stick! Oooh, oooh, yes, yes!

 

[Missing_Linc]

How reliably do humans? I've once received a horrendous gas bill because gas company worker misread 1300 as 1800 (that's in cubic meters)...

I worked in billing for an energy company in the UK (customer service, nothing flash or fun) and human errors in the call centre, just from typing in the incorrect details, occurred often enough that we had posters on the walls reminding people to check their decimal points.

I'd still rather have a human do the work. You can take a human in to a meeting room and admonish them. You can't really do that with an AI.

 

[mwaid1988]

This is stupid. Most gauges and thermometers that need monitoring are IOT connected or LAN connected and you can view and log them on a computer of some kind. Like HOME ASSISTANT. These slop companies are just plain stupid. Even city water meters are read by software. Same for power. They don't have some mofo looking over your back fence with binoculars anymore! They are absolutely psycho and know nothing.

 

[randomname19937]

It's not practical for production environments (where the gauges are likely to be monitored remotely) but makes sense when the inspection is actually an intervention. In that case the inspector might not have access to the data collection system (or the system may even be down).
Finally, the most important part (in my opinion) was when the robot was checking the door visually. There are lots of things that should or should not be happening that are not picked by sensors, but are easy to assess visually.