Google DeepMind and Boston Dynamics bring robotics back to reliable motion

Google Developers has published the Google I/O 2026 video session “Physical AI: the new era of robotics”, a useful marker for where robotics is now being pushed: out of clean demonstrations and into the messier physical world. The session features Jacklyn Dallas, Kanishka Rao and Alberto Rodriguez, and sits alongside the broader Google I/O AI sessions and Dialogues sessions.

The key phrase is “physical AI,” or embodied artificial intelligence. This is not just a new label for robots with cameras. The idea is that models should not only process text, images or code, but also reason about bodies, space, contact, friction, movable objects, shifting human intent and actions that have physical consequences. In software, a bad answer can often be regenerated. In the real world, a bad movement can break an object, injure a person or stop a workflow.

That is why the pairing of Google DeepMind and Boston Dynamics matters. DeepMind brings the weight of models, learning and general reasoning. Boston Dynamics brings years of experience with machines that actually move, balance and operate in changing environments. Even without a concrete product announcement, that combination points to the industry’s current bottleneck: intelligence that looks impressive on a screen has to become reliable enough to control a body.

Most automation so far has been scripted. An industrial robot repeats a precise path. A warehouse system works inside a constrained map. A humanoid demo is often choreographed around known conditions. Physical AI aims at a harder target: a machine that can understand a task, evaluate a scene, plan a movement and adapt when an object is not where it expected it to be.

That is where the serious part begins. “General-purpose” in robotics cannot mean the same thing it means for a chatbot. A robot has to understand what it should not attempt. It has to distinguish a stable grasp from a risky one, a temporary obstruction from a real obstacle, acceptable force from dangerous contact. Physical reasoning is not a decorative layer on top of a model. It is a safety requirement.

The session description uses strong language about a major leap in embodied AI, but the available material should not be stretched beyond what it says. It does not provide performance numbers, benchmark comparisons, robot specifications or a commercial system announcement. Its value is the signal: major AI and robotics organizations are publicly framing the next phase around machines that do more than execute commands. They must understand the physical context in which those commands have consequences.

For the TECH&SPACE audience, that matters more than another viral robot clip. If physical AI holds up, robotics can move from narrow work cells toward more flexible tasks in logistics, manufacturing, home assistance and service environments. If it does not, the field remains stuck with impressive demos that fail when the world stops being orderly. The boundary between those futures will be measured not in stage language, but in the reliability of motion.