Governance around physical AI is becoming increasingly difficult as autonomous AI systems move to robots, sensors, and industrial equipment. The question isn’t just whether an AI agent can complete a task. This is how its actions are tested, monitored, and stopped when interacting with a real system.
Industrial robots already provide a great foundation for that discussion. The International Federation of Robotics has announced that 542,000 industrial robots will be installed worldwide in 2024, more than double the annual level recorded a decade ago. The number of installed units is expected to reach 575,000 in 2025 and exceed 700,000 by 2028.
Market researchers are applying the physical AI label to a broader group of systems, including robotics, edge computing, and autonomous machines. Grand View Research estimates the global physical AI market to be USD 81.64 billion in 2025 and forecast to reach USD 960.38 billion by 2033, although this category varies depending on how vendors define intelligence in physical systems.
From model output to physical action
Governance challenges differ from software-only automation because physical systems can operate around workplaces, infrastructure, and human users. It can also be connected to equipment that requires clear safety limits. The output of the model can be robot movements or machine instructions. Decisions can also be based on sensor data. This makes safety limits and escalation paths part of the system design.
Google DeepMind’s robotics efforts are one recent example of how AI models are being adapted to this environment. The company announced Gemini Robotics and Gemini Robotics-ER in March 2025, describing them as models built on Gemini 2.0 for robotics and embodiment AI. Gemini Robotics is a visual-verbal-behavioral model designed to directly control robots, while Gemini Robotics-ER focuses on embodied reasoning such as spatial understanding and task planning.
Robots using this type of model may need to identify objects, understand instructions, and plan a sequence of actions. You also need to evaluate whether the task was completed correctly. This creates control problems that involve both model behavior and mechanical limitations of the system.
Google DeepMind said that a useful robot must be versatile, interactive, and dexterous. Generality covers unfamiliar objects and environments. Interactivity relates to human input and changes in context. Dexterity refers to physical tasks that require precise movements.
Gemini Robotics can follow natural language instructions and perform multi-step manipulation tasks, Google DeepMind said in a statement. Examples include folding paper, stuffing objects into bags, and handling objects not seen during training.
The technical requirements for physical AI are broader than language understanding. The system requires visual recognition and spatial reasoning. It also requires task planning and success detection. In robotics, success detection is important because the system must determine whether a task is complete, whether it should retry, or whether it should stop.
Google DeepMind’s Gemini Robotics-ER 1.6, introduced in April 2026, shows how these capabilities are packaged into new models. The company explains that the model supports spatial logic, task planning, success detection, and has the ability to reason about intermediate steps and decide whether to proceed or try again.
According to Google’s developer documentation, Gemini Robotics-ER 1.6 is available in preview through the Gemini API. The documentation describes Gemini as a vision language model that brings agent capabilities to robotics. These capabilities include visual interpretation, spatial reasoning, and planning from natural language commands.
Google AI Studio provides a developer environment for working with Gemini models, and the Gemini API provides a route for integrating these models into applications. In the context of embodied AI, tests and prompts are placed closer to the developer building the agent application.
Safety controls are incorporated into the system design
Governance becomes more complex as these systems can invoke tools, generate code, and trigger actions. Controls should define what data the system can access, what tools it can use, what actions require human approval, and how activities are logged for review.
McKinsey’s 2026 AI Trust Study points to the same issues in enterprise AI more broadly. They found that even as AI systems take on more autonomous functions, only about a third of organizations report a maturity level of 3 or higher for strategy, governance, and agent AI governance.
In robotics, safety also includes the physical operation of the machine. Google DeepMind describes robot safety as a multi-layered problem, covering lower-level controls such as collision avoidance, force limitations, and stability, as well as higher-level reasoning about whether a requested action is safe within the context.
The company also introduced ASIMOV, a dataset for evaluating semantic safety in robotics and body-based AI. Google DeepMind said the dataset was designed to test whether the system can understand safety-related instructions and avoid unsafe practices in physical settings.
When a system is connected to robots, sensors, or industrial equipment, it becomes difficult to manage the same controls used for software agents. These include access rights, audit trails, and denial actions. Also includes escalation paths and tests.
Governance frameworks such as the NIST AI Risk Management Framework and ISO/IEC 42001 provide a structure for managing AI risks and responsibilities throughout the system lifecycle. With Physical AI, these controls must take into account the model’s behavior, connected machines, and operating environment.
Google DeepMind is also working with robotics companies as part of its efforts to develop embodied AI. In March 2025, the company announced that it was partnering with Apptronik on humanoid robots using Gemini 2.0 and listed Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools as trusted testers for Gemini Robotics-ER.
The 2026 update also mentioned a collaboration with Boston Dynamics that includes robotic tasks such as instrument reading. This type of use case relies on visual understanding, task planning, and reliable assessment of physical conditions.
Physical AI applies to industrial inspection, manufacturing, and logistics. This also applies to facilities, warehouses, etc. These settings require the system to interpret real-world conditions and operate within defined limits. The governance issue is how to set these limits before autonomous systems can make or execute decisions.
Google DeepMind and Google AI Studio are listed as hackathon technology partners for AI & Big Data Expo North America 2026, which will be held May 18-19 at the San Jose McEnery Convention Center.
(Photo provided by Mitchell Luo)
See: Governance of AI agents comes into focus as regulators flag control gaps
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
