CES has made dancing humanoid robots the most salient benchmark for “physical AI,” even though the hard part starts after the dance moves-getting machines to balance, perceive, plan, and manipulate reliably outside of controlled demos.
The spectacle was easy to find on the show floor: humanoids shadowboxing and dancing and acting out retail scenes, Sharpa’s robotic hand playing table tennis and dealing blackjack. The draw was less about any single robot, and more about a shared claim from the tech stack behind them-that models originally built for chatbots are now being repurposed into control systems that translate sensors into motion.
Nvidia framed that shift as an extension of its broader AI infrastructure push. “The humanoid industry is riding on the work of the AI factories we’re building for other AI stuff,” said Jensen Huang. At CES, the company updated on everything from robot reasoning and planning to partnerships across industrial and consumer brands. The implication for engineering teams is fairly straightforward: platform providers want robotics developers to standardize on their chips, model families, and simulation tooling, so the path from prototype to deployment looks more like shipping software than bespoke mechatronics.
One of the technical reasons the industry is leaning into vision-language approaches is fundamentally practical: robots need to make sense of messy environments. The models can merge camera and sensor input with instruction-following, then feed action policies that handle navigation, grasping, and sequencing-useful in warehouses, light manufacturing, retail operations, agriculture, and healthcare, where McKinsey has estimated general-purpose robotics could reach $370 billion by 2040.
What comes between a CES demo and a reliable shift worker is data and control. Nvidia’s Project GR00T work focuses on synthetic environments and demonstrations of teleoperation to scale training without endless capture of real-world data, including workflows described in new GR00T workflows on environment generation, dexterity, mobility, perception, and whole-body control. That aligns with an industry constraint: there is less data on physical interaction than there is internet text, and humanoids amplify the problem by combining locomotion with manipulation.
Meanwhile, Qualcomm used CES to press a different advantage: power-efficient edge compute with a full robotics stack. The company introduced its Dragonwing IQ10 Series positioning alongside a broader architecture meant to cover everything from household robots to industrial autonomous mobile robots and full-size humanoids. The competitive subtext is a platform land grab; whoever becomes the default “brain of the robot” gets leverage over developer tools, model deployment and fleet operations.
That momentum doesn’t erase the bottleneck that analysts and automation leaders flag again and again: unstructured environments. “Home is very unstructured,” said Jeff Burnstein, president of the Association for Advancing Automation. Unpredictable motion around people, pets, and clutter raises safety and validation requirements that factories can often constrain with layout, guarding, and process design.
Healthcare is emerging as a parallel proving ground, but with narrower task definitions. Nvidia showed off a spine-surgery support robot concept based on specialty arms and sensors rather than a bipedal body. The broader lesson for industry is that “humanoid” is becoming shorthand for a capability goal—generalist operation in human spaces—while many near-term deployments are going to continue to pick whatever embodiment best reduces risk and increases repeatability.