Latest / Minds, Bodies, and Terawatts / Google DeepMind partners with Agile Robots to deploy foundation models on humanoid platform

Transcript

Google DeepMind partners with Agile Robots to deploy foundation models on humanoid platform · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the algorithm gets legs. The Robot Report published on March 26th, 2026, that Google DeepMind has partnered with Agile Robots to deploy its foundation models directly onto Agile's humanoid robot platform. They're bringing Gemini models and Google Cloud infrastructure into a walking, moving body.
• 0:16
  [Guest] This is Google making its play for the physical world. DeepMind has had some of the most impressive robotics research papers for years, from RT-2 to their work on generalist robot policies, but they've lacked a hardware partner to commercialize any of it. Agile Robots gives them that body.
• 0:32
  [Host] Who exactly is Agile Robots? I know the big names in this space, Figure, 1X, Tesla, but Agile isn't one I hear as often.
• 0:40
  [Guest] Agile Robots is a Munich and Beijing-based company founded in 2018, spun out of the German Aerospace Center's robotics lab. They've been focused on precision manipulation and force-sensitive control, the kind of dexterous work that most humanoid companies still struggle with. Their background is surgical robotics and fine motor tasks, which is a very different pedigree than the warehouse-first approach most competitors take.
• 1:01
  [Host] So Google picked a partner known for precision, not brute-force box moving. That seems deliberate.
• 1:06
  [Guest] Completely deliberate. Google's Gemini models are multimodal, they process vision, language, and spatial reasoning simultaneously. Pair that with a platform built for delicate physical interaction and you get something aimed at healthcare, electronics assembly, lab work. The high-value tasks where you can't afford a robot that fumbles.
• 1:21
  [Host] So they're targeting the premium end of the market, not the warehouse floor. Walk me through what "deploying foundation models on a humanoid platform" actually means in practice. What changes for the robot?
• 1:32
  [Guest] Traditional industrial robots run on explicit programming. You code every joint angle, every trajectory, every conditional response. A foundation model replaces that entire stack with something closer to intuition.
• 1:42
  [Host] So instead of programming every specific movement, the robot reasons its way through new situations on its own? That's a fundamental shift in how these machines work.
• 1:51
  [Guest] Right. The robot sees a novel object, understands what it is through Gemini's vision capabilities, reasons about how to grasp it, and executes, all without specific pre-programming. The same robot can work in a hospital pharmacy Tuesday and a semiconductor cleanroom Wednesday without being reprogrammed for each.
• 2:06
  [Host] That flexibility is what makes this commercially interesting. One platform, many applications. And a piece on unscarcity dot ai called "Humanoid Robots in 2025: Costs, Companies, and Timeline" puts hard numbers on why this matters. It reports that Unitree's G1 has already pushed lease rates to roughly two dollars and eighty-seven cents per hour for continuous operation, and that over sixty companies are now racing toward commercial humanoid deployment. The economics are reaching a tipping point.
• 2:31
  [Guest] That article makes a specific argument about where the bottleneck sits. It's not the hardware anymore. Unitree, Fourier Intelligence, Figure, they're all driving chassis costs down fast.
• 2:40
  [Host] So the hardware is getting cheap, but can a two-dollar-an-hour robot actually do high-value work? That seems like the real question.
• 2:47
  [Guest] That's the question. The article identifies the software intelligence layer as the remaining cost barrier, the thing that determines whether a cheap robot can only do simple warehouse picks or whether it can handle a fifty-dollar-an-hour lab task. Google just collapsed that barrier by making DeepMind's foundation models available through a cloud subscription. A robotics company that couldn't afford to build its own vision-language-action model can now lease one.
• 3:10
  [Host] They're showcasing this at the 2026 Robotics Summit and Expo in Boston. Is this a demo or a product?
• 3:16
  [Guest] That's the right question, because robotics has a long history of impressive demos that go nowhere. But the timing matters. Google has been under pressure from investors to show commercial returns on DeepMind's robotics research.
• 3:27
  [Host] And the spending is enormous. Alphabet's 2025 annual report showed over forty-five billion in R&D. Shareholders want to see products, not papers.
• 3:35
  [Guest] This partnership lets Google monetize its AI through someone else's hardware, collecting cloud compute fees and model licensing without building factories. Google gets commercialization without manufacturing risk, and Agile gets the most capable AI models on Earth without building them.
• 3:48
  [Host] So who loses in this arrangement? If you're a robotics company that doesn't have a foundation model partner of this caliber, where does that leave you?
• 3:56
  [Guest] In serious trouble. Figure has OpenAI, Tesla builds in-house, and now Agile has Google DeepMind. If you're a mid-tier humanoid startup without a major AI alliance, you just became an acquisition target or a footnote.
• 4:08
  [Host] It's consolidating into an arms race where the price of entry is a world-class foundation model. And there are maybe five organizations on the planet that can build one.
• 4:17
  [Guest] That's right. Which means the competitive landscape is about to narrow fast. The mid-tier is getting squeezed from every direction.
• 4:24
  [Host] What about the geopolitical dimension here? Agile has operations in Beijing, and Washington is simultaneously tightening AI chip export controls. That seems like a collision waiting to happen.
• 4:33
  [Guest] That's the wrinkle nobody's talking about yet. Agile Robots has dual headquarters, Munich and Beijing, with significant Chinese investor backing. Google partnering with a company that has deep Chinese ties creates a tension that will eventually need resolving. The partnership announcement focuses on the German side of the operation, but the technology transfer questions are going to get loud.
• 4:52
  [Host] Could we see a CFIUS review or export restrictions affecting this deal? It wouldn't be the first time a strategically sensitive tech partnership got scrutinized by regulators.
• 5:01
  [Guest] Not at all. Foundation models capable of controlling physical robots in sensitive environments, that's dual-use technology by any reasonable definition. Whether regulators act depends on how much attention this gets, and a Boston expo with major press coverage tends to generate exactly that kind of attention.
• 5:17
  [Host] Let's zoom out to the industry level. When Google shows up with real partnerships and real infrastructure, every competitor has to respond. What does that do to the pace of commercialization?
• 5:27
  [Guest] It accelerates everything. Figure will likely deepen its OpenAI integration. Tesla will push Optimus harder. Chinese firms like Unitree and UBTECH will seek their own model partnerships, probably with Baidu or Alibaba's AI labs.
• 5:38
  [Host] So the differentiator is no longer the robot's body. It's the quality of the mind running it.
• 5:43
  [Guest] And that mind is now a commodity the cloud giants control. We're entering a phase where the AI layer matters more than the hardware underneath it.
• 5:52
  [Host] That connects to another piece on unscarcity dot ai called "AI Layoffs 2025 to 2030: 1.17 Million Jobs Cut and Counting." That article breaks down how the cuts aren't random. It identifies financial services, tech, and media as the three sectors absorbing the heaviest losses, and it documents that companies announcing AI-driven layoffs saw their stock prices rise an average of four percent within thirty days. The market is rewarding headcount reduction tied to AI investment.
• 6:17
  [Guest] The article makes a critical point about sequencing. Those 1.17 million cuts were concentrated in knowledge work, software, customer service, back-office operations. That was the first phase, AI replacing tasks done on screens.
• 6:28
  [Host] And now the same foundation models are going into physical robots. The displacement frontier just expanded way beyond the cubicle.
• 6:34
  [Guest] The Google-Agile partnership opens that second phase. The same models that replaced analysts and copywriters are now entering physical platforms. What the article tracks as a white-collar phenomenon is about to become a blue-collar and pink-collar phenomenon simultaneously.
• 6:47
  [Host] And the speed here is worth noting. Those knowledge-worker cuts took about eighteen months to accumulate. Physical deployment could move faster because the AI models are already trained. Companies don't need to wait for the intelligence to mature, they just need the body to show up.
• 7:02
  [Guest] The software is ahead of the hardware for the first time. Usually it's the opposite, you build the machine and then spend years writing the control code. Now the control code is world-class and waiting. Every new humanoid platform that rolls off a line can immediately access Gemini-level reasoning through the cloud.
• 7:19
  [Host] This feels like the moment the mind and the body merge into one commercial product. Patrick Déglon's book Unscarcity calls this the convergence, where AI cognition and physical robotics stop being separate industries and become a single technological force. The Google-Agile deal is that convergence happening in a press release. You can explore over a hundred research articles, listen to the audiobook, and read the full book at unscarcity dot ai. This has been Minds, Bodies, and Terawatts.