Physical AI: Governing Autonomous Systems

The rapidly evolving landscape of artificial intelligence is pushing beyond the digital realm and into the physical world, introducing complex governance challenges. As autonomous AI systems are integrated into robots, industrial equipment, and a wider array of sensors, the question shifts from mere task completion to a critical examination of how these agents’ actions are meticulously tested, continuously monitored, and effectively halted when interacting with real-world physical systems.

The industrial robotics sector already serves as a significant proving ground for these discussions. Data from the International Federation of Robotics highlights a surge in adoption, with 542,000 industrial robots installed globally in 2024 – more than doubling the figures from a decade prior. Projections indicate this trend will continue, with installations anticipated to reach 575,000 units in 2025 and surpass 700,000 units by 2028, underscoring the accelerating pace of automation in manufacturing and beyond.

Market intelligence firms are increasingly applying the “Physical AI” moniker to a broad spectrum of technologies, encompassing advanced robotics, edge computing solutions, and autonomous machinery. Grand View Research estimates the global Physical AI market was valued at approximately $81.64 billion in 2025 and forecasts a substantial expansion to $960.38 billion by 2033. However, the precise valuation and scope of this market are inherently tied to how vendors define and implement intelligence within physical systems.

Bridging the Digital Divide: From Model Output to Tangible Action

The governance complexities of Physical AI diverge significantly from those of software-only automation. Physical systems operate within dynamic environments, interacting with workplaces, critical infrastructure, and, crucially, human users. Furthermore, their integration with equipment necessitates stringent safety limitations. A model’s output can directly translate into a robot’s movement, a machine’s operational instruction, or a critical decision informed by real-time sensor data. This direct mapping elevates the importance of embedding safety parameters and defining clear escalation protocols directly into the system’s design architecture.

Google DeepMind’s advancements in robotics offer a compelling illustration of how AI models are being engineered for these complex physical applications. The introduction of Gemini Robotics and Gemini Robotics-ER in March 2025, built upon the Gemini 2.0 architecture, signifies a strategic push into embodied AI. Gemini Robotics is designed as a vision-language-action model capable of direct robot control, while Gemini Robotics-ER is optimized for embodied reasoning, encompassing sophisticated spatial comprehension and intricate task planning capabilities.

A robotic system leveraging these advanced models must possess the ability to not only identify objects and interpret natural language commands but also to meticulously plan a sequence of physical movements. Equally vital is the system’s capacity to accurately assess task completion. This introduces a sophisticated control problem that integrates the predictive behavior of the AI model with the inherent mechanical constraints and operational boundaries of the physical hardware.

Google DeepMind has articulated that truly effective robotic systems require a triad of attributes: generality, interactivity, and dexterity. Generality ensures adaptability to novel objects and unpredictable environments. Interactivity allows for seamless integration with human input and dynamic situational adjustments. Dexterity pertains to the precision and finesse required for complex physical manipulation tasks.

In its unveiling, Google DeepMind highlighted Gemini Robotics’ ability to follow natural language instructions and execute multi-step manipulation tasks, providing examples such as precise paper folding, organized item packing, and handling objects it had not encountered during its training phases. This demonstrates a leap towards more intuitive and versatile robotic interaction.

The technical prerequisites for effective Physical AI extend beyond mere language comprehension. These systems demand robust visual perception capabilities, sophisticated spatial reasoning, and advanced task planning algorithms, crucially complemented by reliable success detection mechanisms. In the context of robotics, success detection is paramount; the system must autonomously determine if a task has been successfully executed, if a retry is warranted, or if the operation should be safely aborted.

Google DeepMind’s Gemini Robotics-ER 1.6, launched in April 2026, exemplifies the evolving packaging of these critical functions into newer model iterations. This version is lauded for its enhanced support for spatial logic, sophisticated task planning, and more refined success detection. It possesses the capability to reason through intermediate steps, making informed decisions on whether to proceed with an action or initiate a re-evaluation.

According to Google’s developer documentation, Gemini Robotics-ER 1.6 is currently accessible in a preview phase through the Gemini API. The documentation positions it as a powerful vision-language model that extends Gemini’s agentic capabilities into the realm of robotics, encompassing visual interpretation, nuanced spatial reasoning, and planning derived from natural language directives.

Google AI Studio provides a dedicated development environment for experimenting with Gemini models, while the Gemini API offers a direct pathway for developers to integrate these powerful models into their bespoke applications. In the burgeoning field of embodied AI, this proximity empowers developers to more effectively test and refine prompts for agentic applications.

Embedding Safety: From Oversight to Foundational Design

The governance framework for AI becomes significantly more intricate when systems are empowered to invoke external tools, generate code, or initiate complex actions. Robust controls are essential to precisely define data access permissions, specify the permissible toolset, delineate which actions necessitate human authorization, and ensure comprehensive activity logging for rigorous review and auditing.

McKinsey’s extensive 2026 AI trust research underscores this pervasive challenge across the enterprise AI landscape. The study revealed that despite the increasing autonomy of AI systems, only approximately one-third of organizations report achieving maturity levels of three or higher in key areas such as AI strategy, governance, and particularly, agentic AI governance.

Within the domain of robotics, safety considerations extend directly to the machine’s physical behavior. Google DeepMind conceptualizes robot safety as a multi-layered challenge, encompassing low-level controls such as sophisticated collision avoidance, precise force limitation, and inherent stability, alongside higher-level reasoning that evaluates the contextual safety of a requested action.

The company has also introduced ASIMOV, a dedicated dataset specifically designed for evaluating semantic safety within robotics and embodied AI. Google DeepMind states that ASIMOV is engineered to rigorously test a system’s capacity to comprehend safety-critical instructions and to proactively avoid hazardous behaviors in real-world physical settings.

The very controls that govern software agents become exponentially more challenging to manage when these systems are interconnected with robots, sensitive sensors, or sophisticated industrial machinery. These include granular access rights, immutable audit trails, and sophisticated refusal behaviors. Furthermore, robust escalation pathways and comprehensive testing protocols are indispensable.

Established governance frameworks, such as the NIST AI Risk Management Framework and ISO/IEC 42001, provide structured methodologies for managing AI risks and responsibilities throughout the entire system lifecycle. In the context of Physical AI, these existing controls must be dynamically adapted to account for the unpredictable behavior of AI models, the interconnected nature of physical machines, and the dynamic complexities of the operating environment.

Google DeepMind has actively engaged with leading robotics companies as part of its commitment to advancing embodied AI development. In March 2025, the company announced a strategic partnership with Apptronik, focusing on the development of humanoid robots powered by Gemini 2.0. It also identified industry leaders such as Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools as trusted testers for Gemini Robotics-ER, signifying a collaborative approach to validation.

The 2026 update also referenced ongoing collaborative efforts with Boston Dynamics, specifically targeting robotics tasks such as sophisticated instrument reading. This advanced use case critically depends on the fusion of advanced visual understanding, intricate task planning, and highly reliable assessment of complex physical conditions.

The application of Physical AI spans critical sectors including industrial inspection, precision manufacturing, and intricate logistics operations. Its utility extends further into the management of vast facilities and complex warehouse environments. These demanding settings require AI systems that can accurately interpret real-world conditions and operate strictly within predefined operational and safety limits. The fundamental governance question that emerges is how these critical limits are meticulously defined and enforced *before* autonomous systems are granted the authority to make or execute decisions.