For years, the robotics industry has operated on a simple, if frustrating, premise: build a robot, then build a bespoke brain for it. A different arm, a new set of wheels, a distinct task? Time to start from scratch. This painstaking, one-off approach has left us with an army of specialists but no true generalists. It’s why your Roomba can’t make you a sandwich and a factory arm can’t walk the dog. But what if one AI could learn to pilot them all?
That’s the audacious goal at Google DeepMind, where Carolina Parada, the head of the robotics team, is overseeing a quiet revolution. In a recent, wide-ranging interview with The Humanoid Hub, Parada laid out a vision that swaps bespoke programming for a universal, adaptable intelligence. The team’s “north star,” she says, is nothing less than “solving AGI in the physical world.” While the rest of the world was mesmerized by ChatGPT’s poetry in 2022, Parada notes her team was less surprised, having worked on large language models internally. The real lesson, she felt, was seeing the immense value of putting research into the hands of the public.
Gemini’s Brain, in a Robot’s Body
The engine driving this ambition is Gemini Robotics 1.5, the latest iteration of DeepMind’s foundational model for embodied AI. This isn’t just another chatbot plumbed into a chassis. It’s a true vision-language-action (VLA) model, designed from the ground up to perceive, reason, and act in the messy, unpredictable physical world. “Gemini Robotics adds the ability to reason about physical spaces – allowing robots to take action in the real world,” as described by Google.
The 1.5 upgrade focuses on three pillars: generalization, interactivity, and dexterity. More importantly, it introduces what DeepMind calls “physical agents.” This system uses a two-part brain:
- Gemini Robotics-ER 1.5: The “Embodied Reasoning” model acts as the strategic planner. It takes a complex command, like “clean up this spill,” and breaks it down into logical steps. It can even use tools like Google Search to look up information it doesn’t have.
- Gemini Robotics 1.5 (VLA): This is the motor cortex, taking the step-by-step plan from the reasoning model and translating it into precise physical actions for whatever body it finds itself in.
This architecture allows the robot to “think before acting,” generating an internal monologue to reason through a problem, making its decisions more transparent and, frankly, more intelligent.
The Holy Grail: Cross-Embodiment Transfer
The most significant leap, however, is what Parada calls “cross-embodiment transfer.” The idea is that a skill learned by one robot can be seamlessly transferred to a completely different machine, without retraining. “It really is the same set of weights that works in all of them,” Parada explains, referring to tests across platforms as different as the bi-arm ALOHA, the Franka robot, and Apptronik’s Apollo humanoid.
This is a radical departure from the industry norm. A task learned by a wheeled robot could, in theory, inform how a humanoid performs a similar action. This is the key to escaping the endless cycle of single-platform development. “We really believe in a future where there will be a really broad range of a very rich ecosystem of many different robot types,” Parada states. “If we’re saying that we want to solve AI in the physical world, to us that means that it has to be smart enough to go embody into any robot.”
This concept builds on DeepMind’s previous work with models like RT-X, which was trained on a massive dataset pooled from 22 different robot types across 33 academic labs. That project demonstrated that co-training on diverse hardware imbued the model with emergent skills and a better understanding of spatial relationships. Gemini Robotics 1.5 appears to be the supercharged evolution of this principle.
A Shifting Timeline
For roboticists, the dream of a machine that can simply watch a human and learn has always been a distant one. “It used to be before that everybody in the team was like, ‘ah, this will happen after my career’,” Parada admits. “And now we’re actually having discussions about like, how far out are we talking? Five years? Are we talking 10 years?”
This acceleration is palpable. While Parada acknowledges that humanoids are an “important form factor” because they are designed for our world, she pushes back against the idea that they are the only form factor that matters. DeepMind’s vision is hardware-agnostic. The intelligence is the product, not the metal shell it occupies.
The ultimate challenge? Our homes. Parada believes the home will be “one of the last frontiers” for robotics, precisely because it is so unstructured and chaotic. A factory floor is predictable; a family kitchen is anything but.
One Brain to Bind Them All
DeepMind’s strategy represents a fundamental bet: that the future of robotics lies not in better hardware, but in a more universal, scalable intelligence. By decoupling the AI “brain” from the robotic “body,” they aim to create a foundation model that can learn from every robot simultaneously, compounding its knowledge across a global fleet of machines.
It’s an approach that could finally break the one-robot, one-brain bottleneck that has constrained the field for decades. We’re not just getting a smarter robot; we’re witnessing the birth of a universal pilot, ready to embody whatever machine we can build. The Jetson’s robot butler, it seems, just took a giant, cross-embodied leap forward.
