In detail
1. It already runs, on cheap hardware
Paige Bailey, AI Developer Relations Lead at Google DeepMind, states the current position plainly: “Gemini already runs on robots.” (36:30)
The hardware in the demo she describes is modest. Stanford’s open-source “Pupper” robot dog can be built from 3D-printed components and runs on a Raspberry Pi. She credits the Gemini APIs: “you can say like, hey, pupper, shake my hand. Hey, Pupper, follow me… Like, all of these things are are possible today just with the Gemini APIs by themselves running and invoking all of the different robotic components on device on a Raspberry Pi.” (37:32)
This is what physical AI means concretely: a general-purpose model taking part in perception and action choices, alongside robot-specific software that executes them.
2. The model decides which components to invoke
Her sentence puts the Gemini APIs in charge of “invoking all of the different robotic components”, but the interview doesn’t specify which parts run locally and which use remote API calls, and her wording doesn’t establish that the model’s weights run on the Pi. The Pupper project’s AI documentation (labeled work in progress) describes one layered setup: model calls handle the conversation, robot software executes tool calls through ROS2, and tricks use prerecorded motions. It is not a record of Bailey’s exact demo.
The general pattern separates responsibilities. Dellaert and Hutchinson’s robotics text describes task plans executed by low-level controllers, and Google DeepMind’s Gemini Robotics introduction describes connecting its reasoning model to roboticists’ existing low-level controllers.
| Responsibility | What it contributes | What still needs checking |
|---|---|---|
| Perception and state | Interprets observations of objects, obstacles and the robot | Sensor quality and whether observations are current |
| Task reasoning | Interprets the request and chooses a goal or action | The intended object, permitted task and available actions |
| Motion execution | Turns the chosen action into movement | Reachability, actuator limits and the particular machine |
| Feedback | Observes whether the intended result happened | A completed physical task, not only a successful API response |
Fictional example: a robot is asked to put a blue block in a tray. If the block misses the tray, the task is incomplete even if the model reports success. When testing a prototype, check what happens when observations are stale, the model is unavailable or an action can’t be executed.
That is why the hardware being commodity matters rather than being a footnote. Bailey expects the pattern to spread on that basis: “We’ll start to see many more kinds of robots built with commodity hardware,” (39:04) citing Raspberry Pi and cheaper alternatives that ship faster. She frames this as a forecast, not a fact.
3. Who is testing it
DeepMind runs a robotics trusted-tester program for the Gemini models and for open robotics models it publishes on Hugging Face. Bailey names the participants as “Enchanted Tools, Boston Dynamics, the Figure Team,” (38:38) describing them as “testing out or using Gemini models in order to operate robotics.”
At the time of the episode, that was a testing program rather than shipped products. It has moved since. Google now offers its Gemini Robotics reasoning model, Gemini Robotics-ER 2, to any developer as a preview in the Gemini API, and Boston Dynamics runs the earlier ER 1.6 in a live inspection feature for Spot. The models that directly drive robot motion are still limited to early-access partners.
4. The application she is most interested in is not a robot dog
Bailey’s own enthusiasm points at laboratories rather than consumer hardware. She cites Periodic Labs as “using AI to to control robotics, but also to to have those robots do really interesting material science work. So they can design experiments, run the experiments, test out, you know, the likelihood of different component parts being successful for the creation of semiconductors.” (39:33)
Her argument for it is about tedium and error rates rather than speed: the work is boring and repetitive, and that is where humans make mistakes. Her conclusion is hedged: “I think the more that AI is incorporated into the scientific process, both for designing experiments, but also executing on them, I think the better off we’re all gonna be.” (41:52)
Why it matters
Robotics used to be gated on control software written for one body and one task. When the controller is a multimodal model that takes instructions in language, more of the behavior can come from a general model, which is part of why cheaper hardware becomes workable. A 3D-printed dog taking spoken commands (Bailey puts the self-build cost at roughly $2,000) is still a demo, but a simpler version of the pattern is now for sale. Hugging Face’s Reachy Mini, a $399 to $499 desktop robot, runs a conversation app in which a language model decides when to move the robot’s head and which gestures to play, and that pipeline can run entirely on a local computer.
Disclosure: Episode 47 was sponsored by Galileo, the Season 3 presenting sponsor (read at 49:58).
Hear it from the guest
“So you can say like, hey, pupper, shake my hand. Hey, Pupper, follow me. Hey, Pupper, do the spider or, like, go swimming. Or, hey, Pupper, tell me a joke. Like, all of these things are possible today just with the Gemini APIs by themselves running and invoking all of the different robotic components on device on a Raspberry Pi.”
“Periodic Labs, which are using AI to control robotics, but also to have those robots do really interesting material science work. So they can design experiments, run the experiments, test out … the likelihood of different component parts being successful for the creation of semiconductors.”
Quotes lightly edited to remove filler words.
From the conversation
This explainer is drawn from these episodes — each carries its full transcript.
Concepts in this explainer
RoboticsHugging FacePhysical AIReasoning ModelsMultimodal AI