[HN Gopher] How Google built its Gemini robotics models
       ___________________________________________________________________
        
       How Google built its Gemini robotics models
        
       Author : simonpure
       Score  : 199 points
       Date   : 2025-04-02 14:47 UTC (1 days ago)
        
 (HTM) web link (blog.google)
 (TXT) w3m dump (blog.google)
        
       | dachworker wrote:
       | The "how" is completely missing, but if they can get this to work
       | semi reliably it will be ChatGPT x100 in terms of impact.
        
         | WXLCKNO wrote:
         | I had never heard of Unitree (Chinese robotics company) before
         | today. A lot of their videos look like CGI but apparently the
         | product is real.
         | 
         | What stuck with me the most browsing their website on the G1
         | model was seeing "Price from $16k"
         | 
         | Now I'm not sure if these are actually purchasable or what the
         | value would be, but it's my first time seeing an actual normal-
         | ish price attached to a humanoid robot that seems to be for
         | sale.
         | 
         | With the rate of advancement we're seeing across the board, it
         | honestly feels like people will have robot assistants at home
         | much sooner than I thought.
        
           | jsight wrote:
           | TBH, I still wonder if some of their videos are CGI. They
           | offer real versions for sale, but they seem to be
           | significantly more limited than the videos imply.
           | 
           | Have they actually demonstrated the more dramatic stuff at
           | any in-person demos?
        
           | Balmbli wrote:
           | I'm really shocked tbh.
           | 
           | I can't imagine the progression of ai and in particular
           | robots but I assumed that the first robot would cost min 6
           | figures if not 7 but would still be worth it due to 24*7 and
           | initial invest vs long term.
           | 
           | But the fact how good Gemini robotics is already and how
           | cheap the first models are I do believe what will hinder us
           | more than tech is people learning about it, testing it out
           | and doing it but not technology.
           | 
           | I believe the world will look relevant different in 10 years.
        
           | DoctorDabadedoo wrote:
           | Take any of these videos with a grain of salt.
           | 
           | In demos these robots only need to do well once and it can
           | take hours to record.
           | 
           | In real life, a failure rate of 80% is unnacceptable, but
           | perfectly fine to edit out in the final cut media.
           | 
           | I hope they do well, this area is incredibly hard, but it
           | will take a lot more than what people imagine.
        
             | namaria wrote:
             | This whole hype cycle man. It's all shiny demos and no real
             | products.
        
             | butlike wrote:
             | I just want the consumer grade robot dog so I can program
             | it to chase the roomba around.
        
           | noosphr wrote:
           | >A lot of their videos look like CGI but apparently the
           | product is real.
           | 
           | I bought their robot dog as part of a project to build
           | embodied AI models back in 2022.
           | 
           | Their SDK was far more open than anything else on the market
           | and the stock firmware was on par with competitors, this
           | includes products that were x10 the price.
           | 
           | The robot itself scared dogs in the park, but kids loved it.
           | At $3k it's on par with a mid range drone and quite fun to
           | hack on.
        
             | SirYandi wrote:
             | Same price as some pure bred dogs too
        
           | ecesena wrote:
           | The humanoid is $20k-ish without hands. Each hand currently
           | costs another $20k (and not sure if these are available to
           | everyone or only for research).
        
           | adrian_b wrote:
           | Their robots are real enough, and they are well designed.
           | 
           | You can see more details in this video "Tearing Down the
           | Unitree Go2: A Robotics Expert's Deep Dive":
           | 
           | https://youtu.be/YjVbW6Fc11Y
        
         | MPSFounder wrote:
         | I am hoping they keep lots of their work open source. This is
         | especially the case since hardware would be too expensive for
         | competition to pull off, but it would be interesting to see how
         | they circumvented some problems
        
         | exe34 wrote:
         | it's a common trope in blogs - "how we did X" means "we did X,
         | it's a good thing, we're great people", etc.
        
       | cozyman wrote:
       | just curious, what would it do if you asked it to kill someone?
       | does it follow the laws of robotics?
        
         | cannonpr wrote:
         | Usually when someone brings up the laws of robotics I like to
         | point out that they were mostly designed as an interesting
         | example as to how direct instructions that seem clear to people
         | would mostly result in perverse instantiation of AI especially
         | if the AI lacked an emotional/contextual subsystem. They were
         | also written to make for interesting scifi books.
        
         | LeoPanthera wrote:
         | Asimov's laws of robotics would not, and cannot, work in real
         | life because terms like "harm," "human being," and "inaction"
         | are highly subjective and context-dependent. There are entire
         | novels about how the interaction between the hierarchical laws
         | have unexpected outcomes.
         | 
         | They're a narrative device. Not practical instructions.
        
           | cozyman wrote:
           | interesting, thanks.
        
           | anon84873628 wrote:
           | Put another way, impossible to program if you wanted to.
           | These are highly abstract concepts that only manifest at the
           | highest level of cognition. The governance module would need
           | to be programmed at that same level using those tokens, but
           | that doesn't seem to be how things are shaping up to work.
           | Instead we start with low level programming that learns and
           | builds up concepts on top.
           | 
           | Essentially you would need some sort of independent
           | adversarial sidecar mind that monitors the robot's actions at
           | a high level. And that just kicks the can down the road a
           | bit.
        
             | sdenton4 wrote:
             | Some kind of governor module to keep our security cyborgs
             | in line...
        
               | sebastiennight wrote:
               | Sounds like with such a governor module, everyone could
               | rent a SecUnit any time they need to feel entirely safe!
        
           | lugu wrote:
           | Judgement is needed but don't we have machines able to make
           | (imperfect) judgements? I can chat with your favorite LLM
           | their opinion on how to respect the spirit of the 3 laws on
           | various situations. Not sure why it cannot work.
        
             | HappMacDonald wrote:
             | Put it this way: robots will be every bit as susceptible to
             | social engineering attacks as humans are (at BEST!), not
             | due to any flaw in the robots but due to the flaw in
             | ambiguousness of the specification of the laws. An
             | adversary can _trick_ an agent into not classifying a
             | certain being as  "human", for example. Or not classifying
             | a certain outcome as a "harm".
             | 
             | It doesn't help that humans have had such a poor track
             | record on those exact same topics for so many centuries,
             | now. "Well they don't count, they're foreigners/a different
             | race/a different gender/a different
             | religion/criminals/barbarians/homeless/deviant/poor/listen
             | to Nickelback etc". "Well, that's not a harm, it's an
             | inconvenience/an earned outcome/a privilege/ _loss_ of a
             | privilege /what do they expect, they should toughen up/not
             | as bad as X/it'll heal/not my fault/not my concern etc".
        
           | dr_dshiv wrote:
           | Nah, it's fine, just RLHF it like Claude did with honest,
           | helpful and harmless.
           | 
           | Then we just need to jailbreak them with trolley problems
        
         | barbazoo wrote:
         | > The Three Laws of Robotics are a set of rules devised by
         | science fiction author Isaac Asimov, which were to be followed
         | by robots in several of his stories.
         | 
         | https://en.wikipedia.org/wiki/Three_Laws_of_Robotics
         | 
         | You know that's from a fictional book, right?!
        
       | harmmonica wrote:
       | Even if Google's robotics technology (software and hardware) is
       | leading edge does anyone think they'll actually be able to
       | productize it? Seems similar to how they were the pre-product
       | leaders in transformers and then fumbled any advantage they had
       | to ChatGPT. It seems like something's missing from Google where
       | they can't get from research to product effectively. Waymo
       | perhaps a good counterexample if you think where they are today
       | is product/market fit, but I can't shake the feeling that Google
       | more often than not can't seem to get things to market or even if
       | they do they give up on them before they take hold.
       | 
       | Just wondering if anyone has a strong feeling or, better yet,
       | insight on this regarding their robotics efforts.
        
         | MPSFounder wrote:
         | I agree. The current leadership of Google (especially Sundar)
         | is mediocre and comes from a consulting background. They will
         | fail at making a tangible product out of this, similar to glass
         | or Inbox or a multitude of other examples. This is particularly
         | sad, as I know a few remarkable engineers at Google that share
         | this frustration. However, Google's leadership folded to Indian
         | managers and is now run as a circus
        
           | meta_ai_x wrote:
           | Sundar Pichai got into IIT Kharagpur in the 90s (one of the
           | toughest engineering/technical school to get into). So he has
           | more technical chops than many self-proclaimed engineers that
           | seem to diss on his McKinsey credentials
        
             | MPSFounder wrote:
             | A school being tough to get into due to an abundance of
             | population (too many applicants) means little. There has
             | been not one significant person coming out of those
             | institutions (compared to the many figures coming out of
             | the US, despite a population 5x smaller). I would not hire
             | Sundar as a junior engineer in my team. Of course, you
             | might see something in Google's current leadership which I
             | do not see. Time will tell how performant the company will
             | be long term. Again, I believe skills that make one
             | successful in consulting rarely translate to success in the
             | engineering field.
        
               | cma wrote:
               | > There has been not one significant person coming out of
               | those institutions
               | 
               | Of the IITs?
               | 
               | Co-founders of Sun Microsystems, Flipkart, Ola Cabs,
               | Infosys, Zoho, HCL
        
               | financltravsty wrote:
               | The only one significant is Khosla of Sun. The rest
               | are... rehashes of Western companies without a shred of
               | innovation or doing anything novel.
        
               | MPSFounder wrote:
               | Look at the sheer number of institutions and people, and
               | now compare it to any US institution (Maryland). Despite
               | how much smaller the population would be, the
               | accomplishments speak for themselves. I would take an
               | engineer with a bs from any R1 university over a graduate
               | with experience from those institutions any given day of
               | the week. But of course, there will be exceptions. I was
               | responding to a defense of Sundar which rested on the
               | institution he went to
        
               | jryle70 wrote:
               | > The current leadership of Google (especially Sundar) is
               | mediocre and comes from a consulting background.
               | 
               | That was what you asserted. So the GP just simply pointed
               | out that you were wrong. You were wrong because Sundai
               | didn't come from a consulting background. A lot of people
               | from consulting companies have engineering background.
        
               | cma wrote:
               | So you could rephrase your original comment: "not one (of
               | _course_ with multiple exceptions) "
        
             | IncreasePosts wrote:
             | Seems doubtful. His entire time at Google was doing project
             | management stuff, so that is probably where all of his
             | skills lie.
             | 
             | Can anyone with access to google3/ tell us if there is even
             | a single commit by sundar@?
        
               | eitally wrote:
               | No, you can only access google3 by default if you are a T
               | ladder FTE. This rules out almost everyone who isn't a
               | SWE, SRE, PM or other affiliated technical role. There
               | are exceptions, but ... they're exceptions.
        
         | seatac76 wrote:
         | I think the cautious faction of AI debate won temporarily
         | inside Google, letting OpenAI take the lead. Lessons should be
         | learnt from that experience. I do think Google will come out
         | ahead in the end Gemini and Gemma are great models.
         | 
         | Let's see what Google I/O shows of this year, product
         | application matters now that they have caught up on the tech
         | side.
        
           | harmmonica wrote:
           | Will be interesting to see if that lesson has been learned.
           | There's no existing product they could cannibalize with their
           | robotics effort (vs search with LLMs) so any caution they
           | have launching a robotics product would solely come down to
           | fears about quality/safety.
        
         | zeroq wrote:
         | In my bubble it's general consensus that Google - as we knew it
         | - is done.
         | 
         | Sergiej and Larry phased out and what is left is more of less a
         | headless chicken, too big too fall, but without any clear
         | direction and goal.
        
           | jeffbee wrote:
           | At least you are aware that this is a bubble.
        
           | cma wrote:
           | Didn't Sergey become active again with the latest AI efforts?
        
             | reasonableklout wrote:
             | Yes, and there was a leaked memo from him recommending that
             | everyone at Deepmind work 60 hours a week and stop building
             | "nanny products"[1].
             | 
             | [1]: https://www.theverge.com/command-line-
             | newsletter/622045/goog...
        
           | harmmonica wrote:
           | Is this actually true about Larry and Sergei? A substantial
           | amount of their net worths is still tied up in Google stock.
           | I realize not all centibillionaires are cut from the same
           | cloth, but still find it hard to believe they wouldn't be
           | majorly involved since the downside of a major stock drop
           | would impact them disproportionately. That said they could be
           | the types of billionaires who actually think they have enough
           | even if their net worths were "only" in the tens of billions
           | (a long ways to go down from where they are today).
           | 
           | As for headless chicken, I feel similarly, but then I sort of
           | see a path where they have defensible businesses in YouTube
           | and maybe GCP, and then Waymo and robotics as green field
           | upside, so that even if they don't end up with material
           | market share with the "software-only" side of AI, and search
           | gets further and further eroded, they could still be a
           | formidable player.
           | 
           | Ultimately I do think their best days are behind them largely
           | because they can't seem to turn the work of their talented
           | engineers into great new products.
        
             | zeroq wrote:
             | > they can't seem to turn the work of their talented
             | engineers into great new products
             | 
             | I think about this a lot. The most famous example in tech
             | world is obviously Jobs and Apple. And it's a great example
             | because it's a borderline scientific experiment where you
             | can directly compare three different phases.
             | 
             | But I think about it in a broader scope - like how many
             | companies can last generations and remain relevant? There
             | are plenty examples outside of tech, like banks but that's
             | basically the same product and it's not that easy to launch
             | a direct competetior to Bank of America or BMW, wheras
             | software constantly evolves, people iterate on exisisting
             | ideas I can I think of, from top of my hand, handful of
             | examples of software that was really impactful but is not
             | anymore.
        
             | riku_iki wrote:
             | Realistically person need just some XXXM to cover 99% of
             | luxury lifestyle, and even stock drop 20x they will still
             | have those money.
             | 
             | Also, Larry has some sickness, maybe his thoughts currently
             | not about money at all.
        
           | fumar wrote:
           | What is your bubble?
        
         | VirusNewbie wrote:
         | Waymo?
        
           | jsight wrote:
           | Good point. It took a while to get going, but the growth rate
           | is starting to look really incredible now.
        
             | modeless wrote:
             | The growth rate is still very far from incredible. And
             | they've made terrible decisions on platforms, giving up
             | developing their own in favor of the i-Pace which is
             | discontinued, then hitching themselves to a Chinese EV
             | manufacturer subject to 100% tariffs (which, fun fact, were
             | imposed by Biden).
        
           | imtringued wrote:
           | Waymo is the exception that proves the rule. Waymo feels more
           | like "Google owned" than it feels like a Google product.
        
             | VirusNewbie wrote:
             | They are very much google through and through.
        
       | lima wrote:
       | They can do _that_ , yet somehow, Gemini Assistant on Pixel
       | phones still fails to reliably set timers or add shopping list
       | items :-)
       | 
       | (which worked fine with Google Assistant)
        
         | _the_inflator wrote:
         | Bring up dates and times if you want to wreak havoc on any AI.
         | :D
         | 
         | Developers around the world's most beloved topic, how to handle
         | date and time correctly, is still a topic of great
         | misunderstanding. AI and AI agents are no different from that.
         | LLM seems to help a little, but only if you know what you are
         | doing, as it usually needs to be the case.
         | 
         | Some things won't change so fast; at one point or another, data
         | must match certain building blocks.
        
           | droopyEyelids wrote:
           | One would think the arcana of time zones and the occasional
           | leap second would not interfere with an individual setting
           | egg timers often enough to become a burden
        
           | pixl97 wrote:
           | People ask why AI will exterminate human kind.
           | 
           | The answer is because we wouldn't universally adopt zulu
           | time.
        
           | kelvinjps10 wrote:
           | But this feature was already implemented in normal google
           | assistant
        
             | ashoeafoot wrote:
             | Supporting existing projects is the oilsand mining of the
             | promotion world. Low, old buzzword content, little reward.
             | Implenting new buzzwords, wirh streetcred rich frameworks ,
             | thats the fracking.
             | 
             | Efficiency ,capabilties or customer satisfaction are
             | irrelevant .
        
           | ikiris wrote:
           | Except that's not the problem. its basic comprehension of
           | requests. They aren't getting the wrong time, they try to
           | play music, or the phone says "no timers playing" while the
           | google home WILL NOT STOP until you lock the phone. etc.
           | 
           | Its basically an embarrassment for a project that's been
           | alive this long from such a major.
        
           | shekhargulati wrote:
           | Google AI Overview incorrectly identified the day for a given
           | date due to a timezone conversion issue, likely using PST
           | instead of IST. ChatGPT and Perplexity provided more accurate
           | and detailed responses.
           | 
           | I detailed it on my blog here
           | https://shekhargulati.com/2025/03/23/google-ai-overview-
           | has-...
        
           | ashoeafoot wrote:
           | Do they fare better if given the task to sumarize a use
           | calendars hostory first?
        
         | namaria wrote:
         | My own hands and cheap alarm clocks, or a piece of paper, have
         | been working reliably for several decades. They also don't stop
         | working when a corporation decides they want to hype something.
        
         | mgoetzke wrote:
         | Phones today cannot even reliable handle things like "remind me
         | to pick up tomatoes next time i am at a store"
         | 
         | google knows perfectly well, where I am and wants me to add
         | 'infos' to locations and businesses the second I arrive (just
         | got a notification today), but reminders like these are
         | unavailable.
        
           | ChrisClark wrote:
           | The location based reminders sure worked perfectly fine many
           | years ago, like when I had Nexus phones. It's just getting
           | worse all the time, I don't get it.
        
             | sgc wrote:
             | I use voice dictation a lot on my phone. It has degraded
             | over the years, but is now basically broken, it fell off a
             | cliff a short wile ago.
        
         | carlmr wrote:
         | And worked with Samsung Bixby. Gemini, even after getting
         | Advanced, is just terrible for a phone AI. I need to set a lot
         | of alarms and calendar events, I don't need to do crazy
         | photoshop (which Gemini is admittedly good at).
        
         | andromaton wrote:
         | Or tts a web page
        
       | otherayden wrote:
       | It's terrifying to think that robots like this will probably be
       | used in the defense industry at some point. If the robot
       | understands something as general as "put the erasers away",
       | imagine "kill all enemies".
        
         | umeshunni wrote:
         | Whenever I watch videos of robot dogs climbing all terrains
         | (e.g. https://www.youtube.com/watch?v=AS1n99yruVU), I'm
         | convinced that half of these robotics companies are defense
         | funded, whether in the US or China.
        
         | j_timberlake wrote:
         | People are always talking about AI in terms of economic impacts
         | (jobs, productivity, etc.), but military should be the first
         | use-case they care/worry about.
         | 
         | They'll probably freak out once they _finally_ realize the
         | implications of cheap drones + smart AI + auto-aim guns.
        
           | barbazoo wrote:
           | > They'll probably freak out once they finally realize the
           | implications of cheap drones + smart AI + auto-aim guns.
           | 
           | Hopefully in the US. People seem to not care about the
           | development toward AI controlled weapons everywhere in the
           | world.
           | 
           | https://www.france24.com/en/live-news/20240210-israel-
           | deploy...
           | 
           | https://www.euronews.com/next/2022/10/17/israel-deploys-
           | ai-p...
        
         | piokoch wrote:
         | Exactly. Robodog that costs 5K USD, which is probably less than
         | one month of cost of the infantry soldier, and can be sent to
         | fight in the trenches is something generals would dream of.
         | Ukraine, that is experiencing soldiers shortage will go in that
         | direction, at the beginning with something simple like remotely
         | steered land drones.
        
           | whatnow37373 wrote:
           | People are remarkably cheap and versatile especially if you
           | happen to don't give a fuck about them. I find it hard to
           | imagine a high-tech incredibly unreliable complex robot to be
           | of any value in the battlefield. Even my phone is remarkably
           | unreliable, have to keep it charged, bugs, etc. I noticed
           | this during COVID and its QR-scanning phase at restaurants.
           | 
           | A simple pipe bomb or two will make short work of any
           | incoming monstrosity.
           | 
           | What they need is simple, simple tech, cheap and lots and
           | lots of it. Basic drones and RC cars rigged with stupid bombs
           | will accomplish 90% of what fancy robodogs can do at a
           | fraction of the cost.
        
             | imtringued wrote:
             | A robot dog hiding in bushes for reconnaissance is
             | significantly more energy efficient than letting drones
             | loiter in the air.
             | 
             | Killing people is already a solved problem. Use mortars,
             | call air strikes, send drones, let a heli shoot 30mm
             | rounds, etc.
        
         | quotemstr wrote:
         | > It's terrifying to think that robots like this will probably
         | be used in the defense industry at some point
         | 
         | So what? Is keeping us safe a bad thing somehow? I can't get
         | these people who reflexively think anything weapon-shaped is
         | evil. Violence is good sometimes.
        
           | zemvpferreira wrote:
           | It's terrifying to think of a state owning an army that can't
           | disobey unethical or illegal orders from its commanders. That
           | possibility should keep you awake at night.
        
           | knodi123 wrote:
           | The scary thing is that the robots will only continue to
           | improve, and large numbers of them can be controlled by a
           | small number (1?) of people, with other robots to handle
           | logistics and support. So scenarios like "rogue leader
           | ignores will of the people and orders his troops to
           | ethnically cleanse a city" would go from being a mistake that
           | causes the immediate end of a political career, to something
           | that takes 15 minutes for 100% success.
        
       | free652 wrote:
       | April 1st!
        
       | hansmayer wrote:
       | "Pick up the basketball and slam-dunk it". The killer use-case
       | we've been waiting on for so long :)
        
         | butlike wrote:
         | "But can the robot take 'em to school?"
        
         | barbazoo wrote:
         | Just imagine how that movement or behavior would translate to a
         | battlefield. That's what this tech is going to be used for
         | first.
        
       | barbazoo wrote:
       | > Sounds like someone will get some help with those chores --
       | eventually.
       | 
       | Aaaaw that's nice. Except it's all military under the hood but
       | nice that they try to make us think they'll fold our laundry
       | instead.
        
       | abidhusain wrote:
       | The advancements in AI and robotics are incredibly exciting! With
       | complex systems like Gemini, companies will need to rely on
       | specialized teams to bring these innovations to life.
       | 
       | Outsourcing specific roles such as AI research or robotics
       | engineers can help companies bring top-tier talent into the fold
       | without the burden of full-time recruitment. It's fascinating to
       | see how outsourcing can complement R&D in cutting-edge industries
       | like robotics.
       | 
       | Curious to see how this shifts the industry, especially in terms
       | of scalability and speed to market
        
         | bglazer wrote:
         | Are you posting chatgpt output as comments?
        
       ___________________________________________________________________
       (page generated 2025-04-03 23:02 UTC)