Managing Physical AI is growing more complex as autonomous AI systems are increasingly embedded in robots, sensors, and industrial machinery. The central concern is no longer simply whether AI agents can carry out tasks, but rather how their actions are tested, overseen, and halted when they interact with real-world systems.
Industrial robotics already forms a substantial foundation for this conversation. According to the International Federation of Robotics, 542,000 industrial robots were deployed globally in 2024—more than twice the annual figure recorded ten years prior. The organization forecasts installations will climb to 575,000 units in 2025 and surpass 700,000 units by 2028.
Market analysts are also using the Physical AI label to describe a broader range of systems, spanning robotics, edge computing, and autonomous machines. Grand View Research valued the global Physical AI market at US$81.64 billion in 2025 and projected it would grow to US$960.38 billion by 2033, though the exact scope of the category hinges on how vendors define intelligence in physical systems.
From model output to physical action
The governance challenge here differs from software-only automation because physical systems operate in and around workplaces, infrastructure, and human users. They may also be linked to equipment that demands well-defined safety boundaries. A model’s output can translate into a robot’s movement, a machine instruction, or a decision driven by sensor data. This means safety boundaries and escalation procedures must be built into the system’s design from the start.
Google DeepMind’s robotics research offers a recent illustration of how AI models are being tailored for this domain. In March 2025, the company unveiled Gemini Robotics and Gemini Robotics-ER, positioning them as models built on Gemini 2.0 for robotics and embodied AI. Gemini Robotics is a vision-language-action model intended to control robots directly, while Gemini Robotics-ER is focused on embodied reasoning, encompassing spatial understanding and task planning.
A robot leveraging such a model may need to recognize an object, interpret an instruction, and map out a sequence of movements. It must also evaluate whether the task was carried out correctly. This introduces a control challenge that spans both model behavior and the mechanical constraints of the hardware.
Google DeepMind stated that effective robots require three qualities: generality, interactivity, and dexterity. Generality refers to handling unfamiliar objects and environments. Interactivity involves responding to human input and adapting to changing conditions. Dexterity covers physical tasks demanding precise movement.
In its launch announcements, Google DeepMind said Gemini Robotics could follow natural-language instructions and execute multi-step manipulation tasks. Demonstrated examples included folding paper, placing items into a bag, and manipulating objects it had not encountered during training.
The technical demands of Physical AI extend well beyond language comprehension. Systems require visual perception, spatial reasoning, task planning, and success detection. In robotics, success detection is critical because the system must determine whether a task is complete, whether it should attempt it again, or whether it should halt.
Google DeepMind’s Gemini Robotics-ER 1.6, released in April 2026, illustrates how these capabilities are being consolidated in newer models. The company describes the model as supporting spatial logic, task planning, and success detection, with the capacity to reason through intermediate steps and decide whether to proceed or retry.
Google’s developer documentation notes that Gemini Robotics-ER 1.6 is available in preview via the Gemini API. It is characterized as a vision-language model that extends Gemini’s agentic capabilities to robotics, including visual interpretation, spatial reasoning, and planning from natural-language commands.
Google AI Studio offers a developer environment for experimenting with Gemini models, while the Gemini API provides a pathway for embedding those models into applications. In the context of embodied AI, this brings testing and prompting closer to the developers building agentic applications.
Safety controls move into system design
Governance grows more intricate when these systems can invoke tools, generate code, or initiate actions. Controls must specify what data the system can access, which tools it can employ, which actions require human authorization, and how activity is recorded for later review.
McKinsey’s 2026 AI trust research highlights the same issue across enterprise AI more broadly. It found that only about one-third of organizations reported maturity levels of three or higher in strategy, governance, and agentic AI governance, even as AI systems assume increasingly autonomous roles.
In robotics, safety also encompasses the machine’s physical behavior. Google DeepMind has framed robot safety as a layered challenge, covering lower-level controls such as collision avoidance, force limits, and stability, as well as higher-level reasoning about whether a requested action is safe given the context.
The company also introduced ASIMOV, a dataset for assessing semantic safety in robotics and embodied AI. Google DeepMind said the dataset was created to evaluate whether systems can comprehend safety-related instructions and refrain from unsafe behavior in physical environments.
The same controls applied to software agents become more difficult to manage when systems are connected to robots, sensors, or industrial equipment. These include access permissions, audit logs, and refusal behavior, along with escalation procedures and testing protocols.
Governance frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 offer structures for managing AI risks and responsibilities throughout the system lifecycle. In Physical AI, these controls must address model behavior, connected machinery, and the operating environment.
Google DeepMind has also collaborated with robotics companies as part of its embodied AI efforts. In March 2025, the company announced a partnership with Apptronik on humanoid robots powered by Gemini 2.0, and named Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools among the trusted testers for Gemini Robotics-ER.
The 2026 update also referenced collaboration with Boston Dynamics on robotics tasks such as instrument reading. These use cases rely on visual understanding, task planning, and dependable assessment of physical conditions.
Physical AI is relevant to industrial inspection, manufacturing, logistics, facilities management, and warehouses. These environments require systems to interpret real-world conditions and operate within established boundaries. The governance question is how those boundaries are defined before autonomous systems are permitted to make or carry out decisions.
Google DeepMind and Google AI Studio are listed as hackathon technology partners for AI & Big Data Expo North America 2026, scheduled for May 18–19 at the San Jose McEnery Convention Center.
(Photo by Mitchell Luo)
See also: AI agent governance takes focus as regulators flag control gaps
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
