On April 14, Boston Dynamics and Google DeepMind announced that Spot, the yellow-colour robot dog that spent most of its life running narrowly scripted routines, is getting an AI brain. Gemini Robotics-ER 1.6 is being folded into Spot and into Boston Dynamics’ Orbit inspection platform, giving both stronger spatial reasoning, autonomous decision-making, and the ability to keep learning inside complex industrial sites.
Spot’s hardware has been good for years, but the intelligence needed to make it useful in a messy factory, without a human driving it frame by frame, is only arriving now. That gap — between a machine that can move and a machine that understands what it is moving through — is exactly what a decades-old field called embodied AI has been trying to close.
What “embodied” actually means
The easy mistake is to think embodied AI just means “AI in a robot body.” The deeper claim by experts working on this field is that a body is not a delivery mechanism for intelligence, but a part of the computation itself.
In their book How the Body Shapes the Way We Think, Rolf Pfeifer, longtime director of the Artificial Intelligence Laboratory at the University of Zurich, and Josh Bongard, computer science professor at the University of Vermont, argue that thought is not independent of the body but is tightly constrained, and at the same time enabled, by it.
Pfeifer has worked extensively on building physical robots to test behaviours that emerge from the body’s design and look for results that tell something about intelligence in general.
Australian roboticist Rodney Brooks, who wrote the foreword for the book and whose own work at MIT in the late 1980s and 1990s made the case even more provocatively, asserting against the dominant symbolic-AI paradigm of the time, that robots didn’t need internal world-models and elaborate planning to act intelligently. A layered “subsumption architecture” of simple reflexes, tightly coupled to sensors and motors, could produce robust, real-time behaviour without anyone programming a representation of the world at all.
You don’t need to build an abstract map of a room if your sensors and body can just react to the room as it actually is, moment to moment.
Practically, this plays out as “morphological computation” — offloading work that would otherwise require a brain onto the physical structure of the body.
A passive-dynamic walker can stroll down a slight incline with no motors or control system at all, purely because the leg geometry and joints are shaped to do the computing. A soft, compliant robot hand can grasp an oddly shaped object without a controller that has explicitly modeled that shape, because the material itself deforms and adapts.
A common thread connecting Pfeifer’s Zurich lab, Brooks’s robots, and today’s humanoids that is intelligence is distributed between brain, body, and environment, not confined to just one part.
Why this matters now, and why it’s hard
The current wave of interest is precisely about discovering how hard that distribution problem is at scale. Unlike the AI chatbots that learn from texts, images, and videos, embodied AI must master gravity and balance, which are profoundly difficult because of the countless physical scenarios a robot must face and the “simulation-to-real gap,” where success in simulation rarely translates perfectly to the real world.
Industry data bears this out starkly. Though the embodied AI market is projected to reach $23 billion by 2030, most humanoid robots still last only about 90 minutes on a charge, and policies that succeed 95% of the time in the lab drop to roughly 60% in the real world. The gap between demo and deployment is the field’s central, unglamorous problem.
Embodied AI vs. neuromorphic AI
In common discussions, people tend to conflate embodied AI with neuromorphic AI as both take biology seriously. But they answer different questions.
Embodied AI is about where intelligence lives. The field views cognition distributed across brain, and body, meaning computational work is not fixed to the brain. It is largely agnostic about what kind of processor sits inside the robot; you could run an embodied robot’s “brain” on an ordinary GPU cluster.
Neuromorphic AI is about how the processor itself is built — it is a hardware and algorithm paradigm that tries to mimic the physical mechanics of biological neurons, most commonly through spiking neural networks (SNNs).
SNNs operate through discrete spikes: each neuron integrates incoming signals and fires only when a threshold is exceeded, which makes them well suited to time-sensitive tasks like motion sensing. Neuromorphic chips are also notably power-efficient, since only the neurons actively spiking consume energy at any given moment.
The two areas in robotics increasingly overlap in practice as a growing body of research on “embodied neuromorphic intelligence” puts spiking, event-driven chips inside physical robots specifically because their low power draw and timely response suit real-world, always-on embodied tasks.
Where evolution comes in
If bodies do computational work, an obvious question follows: how do you design the right body for the task, instead of bolting an AI model onto whatever frame engineers have built?
This is the terrain of evolutionary computation.
Yaochu Jin, an Alexander von Humboldt Professor at Bielefeld University, makes the case that neural control and physical form shouldn’t be designed separately and then combined. Instead, the two have to be developed together, the way biological organisms grow nervous systems and bodies in tandem.
Mr. Jin’s research sits at the intersection of evolutionary algorithms, developmental biology, and computational neuroscience, focused on the interaction between evolution, learning, and artificial systems.
His research focuses on co-evolving nervous systems and morphology, and modelling how feedback from the environment shapes an organism’s sensory distribution.
This matters directly for embodied AI’s central bottleneck: robot bodies are frequently mismatched to the tasks and environments they end up in, and manually re-engineering hardware every time a task changes is slow and expensive.
Evolutionary methods offer a way to search body-and-brain design space automatically, letting simulated populations of robotic forms compete and replicate based on task performance before anything is built.
The unresolved problems
Even as capital and models pour into the field of embodied AI, a fairly consistent list of hard problems keeps surfacing. For instance, the gap between simulation and real-world experience in using robotic systems is still not narrow. Policies trained in simulation that can be run millions of times cheaply degrade sharply when deployed onto real hardware.
Sophisticated reasoning models are often too slow to keep pace with a robot arm or leg that needs to react in milliseconds, forcing a split between heavy “thinking” done off-device and lighter reflexive control on it. Short runtimes due to extensive battery use and vulnerable components keep undercutting otherwise successful pilots.
These systems also lack data. Text and image models were scaled on troves of internet data; embodied systems have no equivalent. Efforts like the Open X-Embodiment dataset and Generalist AI’s GEN-0, pretrained on hundreds of thousands of hours of manipulation data, are attempts to build that missing corpus. But to put these systems in real world situations they need to be trained on at least tens of millions of hours of data.
Embodied AI should be viewed as a systems challenge rather than only a software challenge as safe deployment depends on sensors, hardware robustness, operational design limits, human interaction, cybersecurity, and organisational processes, not algorithms alone. Regulators must start asking for evidence of what counts as proof for allowing learning-enabled robotics systems to be deployed in real world.
All of these problems cannot be solved by a single breakthrough. They’re being chipped away in parallel with better simulators to narrow the simulation-to-real gap, larger shared datasets to address data scarcity, split “brain in the cloud, reflex on the robot” architectures to handle the speed mismatch, and slowly maturing safety frameworks to handle governance.
Evolutionary methods of the kind Mr. Jin champions add another angle to that effort. Instead of only improving how a fixed body learns, they ask whether a differently form factor would be an easier problem to solve in the first place.
Embodied AI systems have come a long way. Boston Dynamics’s humanoid robot Atlas on the football field during FIFA World Cup 2026 shows how robots are able to adapt to uneven surfaces. In a separate event, on America’s Got Talent Season 21, China’s Unitree G1 robots performed alongside Wu Yufei, a 26-year-old Chinese dancer. Spot is now powered by an AI brain and can assist e-commerce delivery agents.
It is starting to become clear that intelligence was never just a matter of better software sitting inside better computers. It was also, at least partly, a matter of testing out new form factors. And these designs are starting to get better with advances in quadruped, bidpedal, and a
