BOSTON, 21 AUG 2026 — Boston Dynamics is integrating Google DeepMind's Gemini Robotics-ER 1.6 into its Spot quadruped and its Orbit inspection platform, in a partnership with Google Cloud aimed at spatial reasoning and autonomous decision-making in industrial settings.
A machine that can already walk anywhere is being given something to work out once it gets there.
What was announced
The partnership names spatial reasoning, autonomous decision-making and continuous learning in complex industrial environments as its objectives. Google Cloud is the third party to it.
Walking was the easy half
Spot has been commercially available for years, and what it does well is move — over rubble, up stairs, through a plant, recovering from a shove. That was the hard research problem of the previous decade and it is finished.
What has limited its usefulness is everything after arrival. A robot that walks a fixed patrol route to photograph the same gauges is just a mobile camera on a schedule. That is useful, but it is a narrow job for a machine with this much mobility. Anything requiring it to notice something unexpected, decide the unexpected thing matters, and do something different as a result has needed a person.
The missing piece is spatial reasoning. It is the difference between just seeing a valve and understanding that this specific valve is leaking, the puddle beneath it is new, and the correct response is to inspect the flange instead of continuing the patrol. None of that is a control problem. It is judgement, and judgement is where the last two years of model progress has actually landed.
Orbit is the less-discussed half of this and possibly the more consequential. It is the software that manages a fleet, schedules routes and holds the record of what was inspected and when, which means it is where an inspection regime becomes auditable. Putting reasoning into the platform, not just the robot, means the capability can be applied to archived data. A model that can review a year of inspection imagery for a pattern nobody thought to look for is a much more powerful tool than a smarter patrol robot.
Industrial inspection is the right first application
Of all the jobs a general-purpose robot might do, inspection is the one where the economics already work.
The work repeats, it happens where people would rather not go, and getting it wrong produces a missed reading rather than an injury. Whatever comes back is already digital anyway, whether that is a photograph, a temperature or a gauge value. And the job requires no gripping, manipulation, or close-quarters work with humans, which are the things robotics still finds hard.
It is also a job where the alternative is expensive and getting more so. Plants that need a trained technician to walk a route twice a day are competing for a shrinking pool of people willing to do it, and that is as true in Johor as in Ohio.
The caveat is that autonomy in inspection shifts the operator's responsibility. A patrol robot that reports everything it saw is a tool. One that decides what is worth reporting has made an editorial judgement about the plant's safety, and somebody has to own the case where it decided wrongly.
The choice of platform is worth a sentence too. A quadruped is an expensive way to move a sensor, and Boston Dynamics has always argued that legs earn their cost in environments built for people — stairs, ladder cages, gantries, doorways with sills. That argument is strongest exactly where inspection matters most, which is old plant rather than new. A facility designed in the last decade can often be covered by fixed instrumentation; a refinery built in the nineteen-eighties cannot, and that is where a machine able to climb pays for itself.
Three companies, three layers
The structure of the partnership is as informative as the capability.
Boston Dynamics owns the body and the years of control engineering behind it, DeepMind owns the model, and Google Cloud owns both the infrastructure it runs on and the pipe between the two. None of the three could have shipped this alone inside a sensible timeframe, which describes most of the robotics industry at the moment.
It also raises the connectivity question that this desk keeps arriving at. A robot depending on a cloud model needs the network to be there, and industrial sites are frequently the places where it is not. We wrote yesterday about TDK putting inference onto an industrial sensor precisely to avoid that dependency. Nothing published says how much of this runs locally, and for anyone evaluating it in a plant with patchy coverage, that is the first question.
What it means for regional operators
Spot is expensive, and the interesting question here is not whether Southeast Asian plants will buy quadrupeds this year.
It is that inspection autonomy is arriving as a software capability on hardware that already exists, and the operators most exposed to the labour constraint it addresses are in this region. Electronics fabs, refineries, palm oil processing and port facilities across Malaysia, Singapore, Indonesia and Thailand all run routine inspection regimes staffed by people who are getting harder to hire and more expensive to retain.
Very few plants here will buy a fleet of robot dogs. What will reach them is this reasoning layer arriving on cheaper platforms — fixed cameras, rail-mounted units, sensors already bolted to the machinery. That should happen within a couple of years, because while the model is expensive to build, it is cheap to copy.
If you run an inspection regime, work out what proportion of it is a person confirming that nothing has changed since yesterday. In most plants that is the bulk of it, and it is precisely the part this technology absorbs.
What we could not establish
What the integration actually does in the field. The announcement describes capabilities rather than tasks, and there is a large distance between a model that reasons about space and a robot that reliably does a job unattended.
We also do not know how much processing happens on the robot versus in the cloud, or what happens when connectivity drops. Other open questions include pricing, availability outside the U.S., what "continuous learning" means for customer data, the status of safety certifications, and whether existing Spot fleets can be upgraded.
What to watch
Watch for a customer describing a route that runs unattended. The line between a demonstration and a product is a robot completing an inspection round with no human reviewing every frame. So far, nobody has published that.
Then watch whether the reasoning capability appears on cheaper hardware. If the same model shows up driving fixed cameras or a rail-mounted inspection unit, the addressable market multiplies and the quadruped becomes the flagship rather than the product.
Finally, watch the liability language. The first contract that specifies who is responsible when an autonomous inspection misses a fault will tell the industry more about how fast this deploys than any capability announcement.
View author profile → · Editorial policy
