Humanoid robots have been promised by tech evangelists for decades. We’ve had extraordinary hardware – Boston Dynamics’ Atlas was doing backflips back in 2018 – yet that promise never translated into robots that truly think, adapt, and operate in messy, real-world settings.
The reason for that was quite simple, according to Demis Hassabis, CEO of DeepMind. “I’ve always felt that the bottleneck in robotics isn’t so much the hardware … but it’s actually the software intelligence that I think is always what’s held robotics back.” That bottleneck has finally begun to crack, as per Hassabis.
Today, the narrative isn’t “robots someday.” It’s that software intelligence is collapsing what once took years of bespoke engineering into months – and in some cases, weeks.
Historically, every robotic task was a separate codebase. Pick up a box? A program for that. Pick up a bottle? Another one. Change the lighting, the background, or the orientation? Start over. There was no transfer, no generalization, no accumulation of skills. That’s why factories could deploy robotic arms for narrowly defined tasks, but you never saw a robot walk into a kitchen and do your dishes the same way.
Also read: Google predicts a 2026 robotics boom, Elon Musk and Altman agree
Foundation models have rewritten that rulebook. By training on massive multimodal data and reasoning about the world instead of reacting to fixed inputs, these models can now transfer competence across tasks. You teach a robot one skill and it generalizes to many others.
This is the reason why robotics timelines are compressing dramatically: years of specialized programming are giving way to rapid fine-tuning of generalist models, which can then be adapted with orders-of-magnitude fewer examples than before.
A clear inflection point has been the launch of Gemini Robotics and its follow-up Gemini Robotics 1.5, foundation models designed to bring AI out of screens and into the physical world, according to DeepMind CEO Demis Hassabis. Unlike older AI systems that output text or predictions, these vision-language-action (VLA) models allow robots to perceive environments, plan multi-step tasks, and execute them autonomously – handling complexity without hand-coded behaviors for every contingency.
Also read: World’s smallest programmable robot: 5 crazy things it can do
More importantly, these models enable robots to “think before acting.” That means they can break complex commands into a sequence of steps – a foundational trait of flexible intelligence.
This isn’t small-scale trickery. It’s the kind of capabilities you’d expect in general intelligence. Which brings us back to Hassabis’s larger point: “In the end AGI needs to be able to do all of those things.” He asserts that embodied intelligence – the ability to sense, plan, and act in the physical world – is central to real AGI, not a fringe application.
The industry used to price robotics as highly specialised engineering. Now, it’s pricing them as embodied AI systems. Tasks that were never economical to automate because of integration costs are suddenly on the table.
The first robots that can reason and act are already here. They have just begun to learn from and improve with each deployment, creating a data flywheel that accelerates future capability gains.
And more importantly, as Hassabis observes, this intelligent physical agency is not just an add-on to AI – it’s a core pillar of what real general intelligence will look like.
Also read: Zenith H1 robot is already working in hotels: Specs, features, price revealed
