Transcription
As Google releases Gemini for robotics, we're looking at the state of the intersection of these two fields more broadly. Welcome back to another AI Daily Brief. Today, we're going to do that thing where we take a bit of contemporary news and use that as a lens to look at a broader set of updates that have happened over the last few weeks. And, as I mentioned, we are talking today about the intersection of AI and Robotics.
Now, the specific catalyst for this conversation is that Google has released a new family of AI models that are specifically designed to drive humanoid robotics. Meaning it's a good time to talk about embodied AI. This is a field that is moving extremely quickly, and a big part of that is driven by the advances in the AI models that actually power the robotics.
It's less than six months since Elon Musk unveiled Tesla's Optimus robot at the big, splashy robotaxi event. And while those robots were visually impressive, it came out in the following days that the robots were largely being controlled remotely, from behind the scenes. And as much as that was fodder for the Elon haters, it also reflected the fact that embodied AI is really hard, hard, especially when it comes to AI models that work for generalized tasks. Humanoid robots have so far required specific training for each action, with the AI models largely helping with edge cases and little deviations. For example, the Optimus robots could easily mix a drink during the demo, likely because they were trained to do that. However, they would have had difficulty if a patron asked to shake their hand without a human controlling them. That's the problem that Google DeepMind's new AI model is trying to solve.
Called Gemini Robotics, the new model is built on top of Gemini 2.0, inheriting Gemini's native multimodal functionality. Meaning that the model can process visual, text, and audio inputs. In their announcement blog post, DeepMind wrote: "To be useful and helpful to people, AI models for robotics need three principal qualities: they have to be general, meaning they're there to adapt to different situations; they have to be interactive, meaning they can understand and respond quickly to instructions or changes in their environment; and they have to be dexterous, meaning they can do the kind of things people generally do with their hands and fingers, like carefully manipulate objects."
DeepMind has actually built a pair of models to drive different parts of the functionality required for generalized robotics. The first is their Advanced Vision Language Action, or VLA, model, which is functionally similar to other multimodal LLMs but includes physical actions as a new mode of output. The second is called Gemini Robotics ER, short for embodied reasoning. The model takes the premise behind reasoning models and applies it to physical environments. As DeepMind put it, the model has "advanced spatial understanding."
Now, as an interesting note, this is similar to the way that the current generation of AI agents are being designed. Agent builders typically use a reasoning model for planning and analysis of the situation and then hand that off to a separate model for execution. Meaning that it's not unreasonable to think of embodied AI as agents with eyes and hands. DeepMind says the Google robotics model "leverages Gemini's world understanding to generalize to novel situations and solve a wide variety of tasks out of the box, including tasks it has never seen before in training." As the model is built on top of an LLM, it has a general understanding of language inputs and can take instruction in natural language. One of the demo videos shows a table with a variety of fruit and containers laid out. The embodied AI receives a voice command and deftly places the banana in the clear container without having any specific training on that task. Google also demonstrated a big step up in fine motor skills, with the embodied AI able to close a Ziploc bag and even make an origami crane. The reasoning model, Google Robotics ER, is added to help increase the robot's ability to plan for novel task execution. DeepMind writes: "Combining spatial reasoning and Gemini's coding abilities, Gemini Robotics ER can instantiate entirely new capabilities on the fly. For example, when shown a coffee mug, the model can assign it an appropriate two-finger grasp for picking it up by the handle and a safe trajectory for approaching it." Functionality from reasoning LLMs also carries over into the real world, meaning the robots can do things like play Tic-Tac-Toe or complete a word puzzle using Scrabble tiles. A key breakthrough here is that this system of models allows robots to move from a narrow range of specific tasks to much more generalized applications.
Kirana Gopala Krishnan, who works on the embodied AI team at DeepMind, posted: "Gemini Robotics is out and is the most advanced VLA in the world. I'm especially blown away by the instruction-following results. It's the first time where I've personally felt that building generic embodied intelligence is within reach, like a robot coming to life." Bloomberg's Mark Gurman pointed out that the implications are for much more than just Google DeepMind. He said: "Artificial intelligence is going to be at the core of everything, and really the ultimate hardware expression of AI is robotics. Being able to understand how a human acts, artificially learn from data, and mimic a human—and that's what a robot is."
Now, Google aren't the only ones that have been working on this form of embodied AI. In early February, Figure AI ditched their partnership with OpenAI to use their own models developed in-house. A few weeks later, we got a look at what these models can do. The demo video showed a pair of robots working together to pack away a grocery delivery. The robots had never seen the items before but were able to reason about where the ketchup bottle should go in the fridge. If one's trying to make direct one-to-one comparisons, some might think that this demo wasn't as impressive as Google's demos from this week, with the robots acting much more slowly, seeming less dexterous, and promising a more limited range of tasks. But on the other hand, Figure AI has their own humanoid design and production, while Google were demonstrating their software on hardware sourced from other companies. Still, both companies seem to be working on the same basic system design of pairing a reasoning model with an execution model. When they dropped the OpenAI deal, Figure AI CEO Brett Adcock said: "We found that to solve embodied AI at scale in the real world, you have to vertically integrate robot AI. We can't outsource AI for the same reason we can't outsource our hardware." And Figure AI has begun deploying their robots in real-world settings. They have one pilot program currently underway in the BMW manufacturing plant in South Carolina and a second undisclosed contract that the company says could potentially allow them to reach 100,000 robots shipped. The company indeed showed a video of robots sorting parcels, making many think that the client is one of the large US shipping companies. These are both commercial clients, but much of the excitement and appetite, at least from an investor perspective, is what seems to many as the inevitable future of bringing humanoids into the household setting. Figure AI also seems to have demonstrated that humanoid companies are past the speculative phase, at least in terms of their valuations. Last February, during their Series B, the company was valued at a very decent 2.6 billion, but last month Bloomberg reported that they are in talks to raise their Series C at a valuation of $39.5 billion.
Of course, we are now also living in the world of Deep Seek and Manis, and everyone is wondering what's going on in China. It feels like every day on X you can see a video of some Chinese-produced robot carrying out some feat of dexterity. Earlier this month, one company called X Robot went viral with an extremely lifelike female robot with a good voice model behind it. Now, this video that you're watching here had the sci-fi factor turned all the way up, so who knows how real the product is. Then again, with what we've seen out of Chinese AI in recent months, I certainly wouldn't count it out. One Chinese company that is definitely producing real products is Unitree. They had a huge range of robots and assorted form factors on display at CES in January. You also might have seen the company's latest viral video showing a Kung Fu robot kicking a stick out of a person's hand. Now, many of the videos from trade shows still have a human operator in control, which gets us exactly back to why potentially this Google model is such important news. As Google may have just demonstrated a path to fill in the blanks where Chinese embodied AI is lacking right now. Unitree is offering these G1 units starting at $16,000, but you have to think those prices are going to come down precipitously in the years ahead.
Another key player in embodied AI that's worth mentioning in this roundup is NVIDIA. The chipmaker isn't working on robots per se, but they've definitely made some big advancements in the AI used to train them. In January, Nvidia released their Cosmos World Foundation model. The generative model can be used to create virtual simulations of real-world scenarios for robot training. Improvements in world models have been one of the big breakthroughs over the past few months, with several startups showing off their own versions of the tech in development. The idea is that a digital twin of a robot can be placed in a simulation, which allows synthetic training data to be quickly generated. This doesn't help necessarily with the reasoning and generalization problem that Google is working on, but it does allow for big improvement movements in dexterity and specific movement training. The Cosmos reveal in January also came with some very bullish statements from Nvidia CEO Jensen Huang. He said the "ChatGPT moment for general robotics is just around the corner." He also delivered his keynote address standing in front of a chart showing the AI sector going exponential. After agentic AI, the wave that we're currently in the middle of, the chart spiked even higher for physical AI, consisting of self-driving cars and general robotics. During the speech, Huang said that self-driving cars would likely be the "first multi-trillion dollar robotics industry." And while at this point we haven't seen anything that looks close to a fully capable general-purpose humanoid, Huang did mention that he expects Nvidia's products to power a billion humanoid robots over the coming years.
So far, I've hit a lot of the biggies, but even beyond these companies, VCs are definitely sitting up and paying attention to the potential inflection point we're hitting with embodied AI. Earlier this week, Dexterity Inc. raised $95 million at a $1.65 billion valuation to build robots capable of humanlike dexterity. The company's pitch is remarkably similar to the way Google described their criteria for generalized robotics. CEO Samir Menon described that his robots can touch and recognize objects, are aware of and respond appropriately to surroundings, and will move gracefully and adjust as needed. He added: "The combination of those three is what we engineer and what we believe will drive the future of physical AI." Reza Zadeh, a partner at Lightspeed Ventures, said he was investing more money in the company because he believes we're reaching an inflection point for physical AI. Also last month, a startup called Apptronik raised $350 million in Series A funding at an undisclosed valuation. The company is a spinout from the University of Texas and has been working on humanoid robots for over a decade. The round included participation from Google, with DeepMind partnering with the company to provide the AI to drive their robots. In fact, you could see the Apptronik robots putting Google's embodied AI through its paces in the demo videos from this week. The raise was vastly more money than the $28 million the company had raised prior to this round, and CEO Jeff Cardenas commented that the mega round was necessary because his robots are almost production-ready. He said: "What 2025 is about for Apptronik and the humanoid industry is really demonstrating useful work in these applications with these initial early adopters and customers, and then true commercialization and scaling happening in 2026 and beyond." Explaining the Google partnership, Cardenas said: "It made far more sense than creating their own models. Adding, 'We believe that right now Google is at the top of the game and building some of the best models in the world.'"
So friends, that is a quick update on the state of embodied AI, the intersection of AI and Robotics, and that is where we will wrap today's episode. Appreciate you listening, as always, and until next time, peace.